[verified] feat: complete Ralph task workstreams

packages/node/meshnet_node/activation_compression.py

@@ -0,0 +1,142 @@
+"""Policy-driven zstd compression for activation seam bodies.
+
+Policies are intentionally local to a hop condition: a LAN prefill can favour
+wire savings while a one-token decode keeps a larger raw fast path.  Environment
+overrides make a trace-tuned rollout possible without changing the wire format.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import os
+import time
+
+
+@dataclass(frozen=True)
+class CompressionPolicy:
+    """The measurable conditions required before an activation is compressed."""
+
+    min_input_bytes: int
+    min_savings_bytes: int = 4096
+    min_savings_ratio: float = 0.05
+    level: int = 1
+    enabled: bool = True
+
+
+@dataclass(frozen=True)
+class CompressionResult:
+    body: bytes
+    encoding: str | None
+    input_bytes: int
+    output_bytes: int
+    elapsed_seconds: float
+    decision: str
+
+    @property
+    def compressed(self) -> bool:
+        return self.encoding == "zstd"
+
+
+_DEFAULTS: dict[tuple[str, str], CompressionPolicy] = {
+    # Decode activations usually contain one position; keep that hot path raw.
+    ("lan", "prefill"): CompressionPolicy(64 * 1024),
+    ("lan", "decode"): CompressionPolicy(128 * 1024),
+    ("relay", "prefill"): CompressionPolicy(32 * 1024),
+    ("relay", "decode"): CompressionPolicy(128 * 1024),
+    # The deterministic benchmark can explicitly model either policy family.
+    ("benchmark", "prefill"): CompressionPolicy(64 * 1024),
+    ("benchmark", "decode"): CompressionPolicy(128 * 1024),
+}
+
+
+class CompressionPolicies:
+    """Explicit policies for LAN, relay, and benchmark prefill/decode seams.
+
+    Set ``MESHNET_COMPRESSION_<ROUTE>_<PHASE>_MIN_INPUT_BYTES``,
+    ``..._MIN_SAVINGS_BYTES``, ``..._MIN_SAVINGS_RATIO``, or ``..._ENABLED`` to
+    tune a condition from production traces.  E.g.
+    ``MESHNET_COMPRESSION_RELAY_PREFILL_MIN_INPUT_BYTES=32768``.
+    """
+
+    def __init__(self, policies: dict[tuple[str, str], CompressionPolicy] | None = None) -> None:
+        self._policies = dict(_DEFAULTS if policies is None else policies)
+
+    def for_condition(self, route: str, phase: str) -> CompressionPolicy:
+        key = (route.lower(), phase.lower())
+        try:
+            policy = self._policies[key]
+        except KeyError as exc:
+            raise ValueError(f"unknown compression condition {route}/{phase}") from exc
+        prefix = f"MESHNET_COMPRESSION_{key[0].upper()}_{key[1].upper()}_"
+        return CompressionPolicy(
+            min_input_bytes=_env_int(prefix + "MIN_INPUT_BYTES", policy.min_input_bytes),
+            min_savings_bytes=_env_int(prefix + "MIN_SAVINGS_BYTES", policy.min_savings_bytes),
+            min_savings_ratio=_env_float(prefix + "MIN_SAVINGS_RATIO", policy.min_savings_ratio),
+            level=_env_int(prefix + "LEVEL", policy.level),
+            enabled=_env_bool(prefix + "ENABLED", policy.enabled),
+        )
+
+
+def compress_activation(body: bytes, policy: CompressionPolicy) -> CompressionResult:
+    """Compress only when zstd clears both configured savings thresholds."""
+    started = time.monotonic()
+    if not policy.enabled:
+        return _raw(body, started, "disabled")
+    if len(body) < policy.min_input_bytes:
+        return _raw(body, started, "below_min_input")
+    try:
+        import zstandard as zstd
+
+        candidate = zstd.ZstdCompressor(level=policy.level).compress(body)
+    except Exception:
+        # Compression is an optional transport optimisation, never a reason to
+        # reject an otherwise valid activation.
+        return _raw(body, started, "unavailable")
+    saved = len(body) - len(candidate)
+    ratio = saved / max(1, len(body))
+    if saved < policy.min_savings_bytes or ratio < policy.min_savings_ratio:
+        return _raw(body, started, "below_savings")
+    return CompressionResult(candidate, "zstd", len(body), len(candidate), time.monotonic() - started, "compressed")
+
+
+def decompress_activation(body: bytes, encoding: str | None) -> CompressionResult:
+    """Decode a modern zstd body or preserve a legacy raw body with metrics."""
+    started = time.monotonic()
+    if not encoding:
+        return CompressionResult(body, None, len(body), len(body), time.monotonic() - started, "legacy_raw")
+    if encoding != "zstd":
+        raise ValueError("unsupported X-Meshnet-Encoding")
+    try:
+        import zstandard as zstd
+    except ImportError as exc:
+        raise ValueError("zstd support is unavailable") from exc
+    try:
+        raw = zstd.ZstdDecompressor().decompress(body)
+    except zstd.ZstdError as exc:
+        raise ValueError("invalid zstd activation body") from exc
+    return CompressionResult(raw, "zstd", len(body), len(raw), time.monotonic() - started, "decompressed")
+
+
+def _raw(body: bytes, started: float, decision: str) -> CompressionResult:
+    return CompressionResult(body, None, len(body), len(body), time.monotonic() - started, decision)
+
+
+def _env_int(name: str, default: int) -> int:
+    try:
+        return max(0, int(os.getenv(name, str(default))))
+    except ValueError:
+        return default
+
+
+def _env_float(name: str, default: float) -> float:
+    try:
+        return max(0.0, float(os.getenv(name, str(default))))
+    except ValueError:
+        return default
+
+
+def _env_bool(name: str, default: bool) -> bool:
+    value = os.getenv(name)
+    if value is None:
+        return default
+    return value.strip().lower() not in {"0", "false", "no", "off"}

packages/node/meshnet_node/admission.py

@@ -0,0 +1,225 @@
+"""Fail-closed admission: no routable registration without a fresh matching proof.
+
+This module does not *produce* proof — `doctor` does that, by pushing a bounded
+real forward through the selected shard (NCA-002). This module *decides whether a
+proof covers what is about to be advertised*, and startup calls it immediately
+before it registers with the tracker.
+
+A capability report proves one combination: model artifact, shard range, recipe,
+backend and device. Reusing it for anything else is the exact hole this closes —
+a report that failed, aged out, or describes a different model, shard, recipe or
+device is rejected here, and the node exits without ever registering an endpoint.
+
+Nothing in here branches on a model, vendor or kernel name: identity fields are
+opaque labels that are compared, never interpreted.
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass
+from typing import Any, Callable
+
+from .capability import CapabilityReport
+from .doctor import DoctorSelection
+from .recipe_manifest import Recipe, RecipeManifest
+
+# How long a passing report stays usable. Startup normally validates in-process
+# (age ≈ 0); this bounds how far a report written by an earlier `doctor` run can
+# be carried forward, after which the hardware, drivers or weights may have moved.
+DEFAULT_MAX_REPORT_AGE_SECONDS = 900.0
+
+# A report timestamped this far in the future is not fresh, it is wrong.
+_MAX_CLOCK_SKEW_SECONDS = 60.0
+
+REASON_NO_REPORT = "no-report"
+REASON_NOT_PASSED = "not-passed"
+REASON_STALE = "stale"
+REASON_MODEL_MISMATCH = "model-mismatch"
+REASON_SHARD_MISMATCH = "shard-mismatch"
+REASON_RECIPE_MISMATCH = "recipe-mismatch"
+REASON_BACKEND_MISMATCH = "backend-mismatch"
+
+
+class CapabilityAdmissionError(RuntimeError):
+    """This node may not advertise the selection: the proof does not cover it."""
+
+    def __init__(self, reason: str, message: str) -> None:
+        super().__init__(message)
+        self.reason = reason
+
+
+@dataclass(frozen=True)
+class CapabilityContext:
+    """What is about to be advertised, and the loaded backend that would serve it."""
+
+    backend: Any
+    selection: DoctorSelection
+    recipe: Recipe
+    manifest: RecipeManifest
+    device: str
+
+
+# A validator turns the context into the report the gate then judges. Production
+# uses `probe_capability`; tests pass an explicit test-safe one (see
+# `meshnet_node.testing`) rather than switching this module into a lenient mode.
+CapabilityValidator = Callable[[CapabilityContext], CapabilityReport]
+
+
+@dataclass(frozen=True)
+class AdmissionRequirement:
+    """The one capability a report must prove for this node to register."""
+
+    model_id: str
+    shard_start: int
+    shard_end: int
+    recipe_id: str
+    recipe_version: str
+    backend_id: str
+    device: str
+    max_age_seconds: float = DEFAULT_MAX_REPORT_AGE_SECONDS
+
+    @classmethod
+    def for_context(
+        cls,
+        context: CapabilityContext,
+        *,
+        max_age_seconds: float = DEFAULT_MAX_REPORT_AGE_SECONDS,
+    ) -> AdmissionRequirement:
+        return cls(
+            model_id=context.selection.model_id,
+            shard_start=context.selection.shard_start,
+            shard_end=context.selection.shard_end,
+            recipe_id=context.recipe.id,
+            recipe_version=context.recipe.version,
+            backend_id=context.recipe.backend_id,
+            device=context.device,
+            max_age_seconds=max_age_seconds,
+        )
+
+    @property
+    def shard_label(self) -> str:
+        return f"layers {self.shard_start}–{self.shard_end}"
+
+
+def admit(
+    requirement: AdmissionRequirement,
+    report: CapabilityReport | None,
+    *,
+    now: float | None = None,
+) -> CapabilityReport:
+    """Return `report` if it admits `requirement`; otherwise refuse to register.
+
+    Checks run selection-first, so the operator is told the report is about the
+    wrong thing before being told it is old.
+    """
+    if report is None:
+        raise CapabilityAdmissionError(
+            REASON_NO_REPORT,
+            f"no capability report for {requirement.model_id} "
+            f"{requirement.shard_label}: this node has not proven it can serve it",
+        )
+
+    if report.model.model_id != requirement.model_id:
+        raise _mismatch(
+            REASON_MODEL_MISMATCH,
+            requirement,
+            "model",
+            report.model.model_id,
+            requirement.model_id,
+        )
+
+    if (report.shard.start, report.shard.end) != (
+        requirement.shard_start,
+        requirement.shard_end,
+    ):
+        raise _mismatch(
+            REASON_SHARD_MISMATCH,
+            requirement,
+            "shard",
+            f"layers {report.shard.start}–{report.shard.end}",
+            requirement.shard_label,
+        )
+
+    if (report.recipe.recipe_id, report.recipe.recipe_version) != (
+        requirement.recipe_id,
+        requirement.recipe_version,
+    ):
+        raise _mismatch(
+            REASON_RECIPE_MISMATCH,
+            requirement,
+            "recipe",
+            f"{report.recipe.recipe_id} (v{report.recipe.recipe_version})",
+            f"{requirement.recipe_id} (v{requirement.recipe_version})",
+        )
+
+    if (report.backend.backend_id, report.backend.device) != (
+        requirement.backend_id,
+        requirement.device,
+    ):
+        raise _mismatch(
+            REASON_BACKEND_MISMATCH,
+            requirement,
+            "backend",
+            f"{report.backend.backend_id} on {report.backend.device}",
+            f"{requirement.backend_id} on {requirement.device}",
+        )
+
+    if not report.passed:
+        raise CapabilityAdmissionError(
+            REASON_NOT_PASSED,
+            f"capability validation {report.status} for {requirement.model_id} "
+            f"{requirement.shard_label} with recipe {requirement.recipe_id}"
+            + _diagnostics_suffix(report),
+        )
+
+    now = time.time() if now is None else now
+    age = now - report.validated_at
+    if age > requirement.max_age_seconds:
+        raise CapabilityAdmissionError(
+            REASON_STALE,
+            f"capability report for {requirement.model_id} {requirement.shard_label} "
+            f"is {age / 60:.0f} min old (limit "
+            f"{requirement.max_age_seconds / 60:.0f} min); re-run `meshnet-node doctor`",
+        )
+    if age < -_MAX_CLOCK_SKEW_SECONDS:
+        raise CapabilityAdmissionError(
+            REASON_STALE,
+            f"capability report for {requirement.model_id} {requirement.shard_label} "
+            f"is timestamped {-age:.0f}s in the future; check this host's clock",
+        )
+
+    return report
+
+
+def _mismatch(
+    reason: str,
+    requirement: AdmissionRequirement,
+    field_name: str,
+    reported: str,
+    required: str,
+) -> CapabilityAdmissionError:
+    return CapabilityAdmissionError(
+        reason,
+        f"capability report proves a different {field_name}: it validated "
+        f"{reported}, but this node would serve {required}. A report is only "
+        "proof for the exact combination it ran.",
+    )
+
+
+def _diagnostics_suffix(report: CapabilityReport) -> str:
+    if not report.diagnostics:
+        return ""
+    return " — " + " ".join(report.diagnostics)
+
+
+def probe_capability(context: CapabilityContext) -> CapabilityReport:
+    """Production validator: one bounded real forward through the loaded shard."""
+    from .doctor import validate_loaded_backend
+
+    return validate_loaded_backend(
+        context.backend,
+        context.selection,
+        context.recipe,
+        context.manifest,
+    ).report

packages/node/meshnet_node/cli.py

@@ -237,6 +237,85 @@ def _cmd_config(args) -> int:
     return 0
 
 
+def _doctor_overrides(args) -> dict:
+    """CLI flags that change *what* doctor validates, applied on top of config."""
+    overrides: dict = {}
+    model_name, hf_repo = _resolve_model_flags(
+        getattr(args, "model", None), getattr(args, "model_id", None)
+    )
+    if model_name is not None:
+        overrides["model_name"] = model_name
+        overrides["model_hf_repo"] = hf_repo or ""
+    for flag, key in (
+        ("quantization", "quantization"),
+        ("download_dir", "download_dir"),
+        ("shard_start", "shard_start"),
+        ("shard_end", "shard_end"),
+    ):
+        value = getattr(args, flag, None)
+        if value is not None:
+            overrides[key] = value
+    if getattr(args, "cpu", False):
+        overrides["force_cpu"] = True
+    return overrides
+
+
+def _cmd_doctor(args) -> int:
+    """Validate the selected model/shard with a bounded real forward."""
+    import json
+    import traceback
+
+    from .config import DEFAULTS, load_config, merge_cli_overrides
+    from .doctor import (
+        DoctorError,
+        default_report_path,
+        render_result,
+        resolve_selection,
+        run_doctor,
+        write_reports,
+    )
+
+    debug = bool(getattr(args, "debug", False))
+    cfg = load_config() or dict(DEFAULTS)
+    overrides = _doctor_overrides(args)
+    if overrides:
+        cfg = merge_cli_overrides(cfg, **overrides)
+
+    try:
+        selection = resolve_selection(cfg)
+        result = run_doctor(
+            selection,
+            recipe_id=args.recipe,
+            all_recipes=args.all_recipes,
+        )
+    except DoctorError as exc:
+        # Bad input (no model, unknown recipe): there is nothing to report on.
+        if debug:
+            traceback.print_exc()
+        print(f"ERROR: {exc}", file=sys.stderr, flush=True)
+        if exc.hint:
+            print(f"  {exc.hint}", file=sys.stderr, flush=True)
+        return 1
+
+    written = write_reports(
+        result.reports,
+        Path(args.report) if args.report else default_report_path(),
+    )
+
+    if args.json:
+        print(json.dumps([r.to_dict() for r in result.reports], indent=2, sort_keys=True))
+    else:
+        print(render_result(result, report_path=written))
+
+    if debug:
+        for item in result.results:
+            if item.error is not None:
+                traceback.print_exception(
+                    type(item.error), item.error, item.error.__traceback__
+                )
+    return result.exit_code
+
+
 def _cmd_start(args) -> int:
     """Legacy `start` subcommand — preserves backward compatibility with existing tests."""
     from .config import DEFAULTS
@@ -322,6 +401,7 @@ def main() -> None:
             "  models          List supported models\n"
             "  models --browse Browse HuggingFace Hub\n"
             "  config          Show current config\n"
+            "  doctor          Check this node can really run its selected shard\n"
         ),
     )
 
@@ -367,6 +447,40 @@ def main() -> None:
     # config subcommand
     subparsers.add_parser("config", help="Show current saved config")
 
+    # doctor subcommand — validate the selected shard with a real forward
+    doctor_cmd = subparsers.add_parser(
+        "doctor",
+        help="Check this node can really run its selected model shard",
+    )
+    # These mirror the top-level selection flags. argparse.SUPPRESS keeps an
+    # unpassed subcommand flag from overwriting the top-level one, so both
+    # `meshnet-node --model X doctor` and `meshnet-node doctor --model X` work.
+    doctor_cmd.add_argument("--model", metavar="MODEL", default=argparse.SUPPRESS,
+                            help="Model name or HuggingFace repo ID to validate")
+    doctor_cmd.add_argument("--model-id", metavar="MODEL", default=argparse.SUPPRESS,
+                            help="Alias for --model")
+    doctor_cmd.add_argument("--quantization", "-q", default=argparse.SUPPRESS,
+                            choices=["bf16", "int8", "nf4", "bfloat16", "auto"],
+                            help="Quantization level to validate")
+    doctor_cmd.add_argument("--download-dir", metavar="PATH", default=argparse.SUPPRESS,
+                            help="Model download directory")
+    doctor_cmd.add_argument("--shard-start", type=int, metavar="N", default=argparse.SUPPRESS,
+                            help="Pin shard start layer")
+    doctor_cmd.add_argument("--shard-end", type=int, metavar="N", default=argparse.SUPPRESS,
+                            help="Pin shard end layer")
+    doctor_cmd.add_argument("--cpu", action="store_true", default=argparse.SUPPRESS,
+                            help="Validate CPU execution even when a GPU is available")
+    doctor_cmd.add_argument("--debug", action="store_true", default=argparse.SUPPRESS,
+                            help="Print the full traceback behind a failure")
+    doctor_cmd.add_argument("--recipe", metavar="ID", default=None,
+                            help="Recipe to validate (default: baseline)")
+    doctor_cmd.add_argument("--all-recipes", action="store_true",
+                            help="Validate every recipe in the catalogue, not just the selected one")
+    doctor_cmd.add_argument("--report", metavar="PATH", default=None,
+                            help="Where to write the capability report JSON")
+    doctor_cmd.add_argument("--json", action="store_true",
+                            help="Print the capability report JSON instead of a summary")
+
     # start subcommand (legacy / backward-compat)
     start_cmd = subparsers.add_parser("start", help="Start node (legacy flags)")
     start_cmd.add_argument("--tracker")
@@ -406,6 +520,8 @@ def main() -> None:
         sys.exit(_cmd_models(args))
     elif args.command == "config":
         sys.exit(_cmd_config(args))
+    elif args.command == "doctor":
+        sys.exit(_cmd_doctor(args))
     elif args.command == "start":
         sys.exit(_cmd_start(args))
     else:

packages/node/meshnet_node/doctor.py

@@ -0,0 +1,633 @@
+"""`meshnet-node doctor` — prove the selected shard actually runs.
+
+The doctor answers one question: *would the model/shard/recipe this node is
+configured to serve really execute here?* It answers it the only way that is
+not a guess — by loading the selection through the production backend path and
+pushing a bounded, real forward through the selected layers. Generic hardware
+probing (is there a GPU, can Torch allocate a tensor) proves nothing about a
+shard and is deliberately not what this reports on.
+
+Two shapes of probe, chosen by where the shard sits, never by which model it is:
+
+* head shard — tokenize a short prompt, embed it, run this shard's layers.
+* mid/tail shard — synthesize a small hidden-state tensor in the same wire
+  format peers send, and push it through `forward_bytes`. A tail shard decodes
+  it, which also exercises the final norm and `lm_head`.
+
+Everything here is model-agnostic: `model_id` is opaque, and no vendor or kernel
+name is a branch. Failures are reported as a category plus an actionable hint
+(never a raw traceback, unless the caller asks for one) and produce a *failed*
+capability report — a failure is evidence too, and NCA-003 refuses to register
+without a fresh passing one.
+"""
+
+from __future__ import annotations
+
+import base64
+import struct
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Callable, Mapping, Sequence
+
+from .capability import (
+    STATUS_FAILED,
+    STATUS_PASSED,
+    CapabilityReport,
+    build_capability_report,
+)
+from .recipe_manifest import (
+    DEFAULT_RECIPE_ID,
+    Recipe,
+    RecipeManifest,
+    RecipeManifestError,
+    load_recipe_manifest,
+)
+
+# The probe is deliberately tiny: enough tokens to drive every layer in the
+# shard once, small enough that `doctor` costs seconds beyond the model load.
+PROBE_TOKENS = 4
+PROBE_PROMPT = "meshnet capability probe"
+
+# Failure categories. These are what an operator acts on, so they name the thing
+# to fix, not the exception that surfaced it.
+CATEGORY_NO_MODEL = "no-model-selected"
+CATEGORY_MISSING_DEPENDENCY = "missing-dependency"
+CATEGORY_MODEL_UNAVAILABLE = "model-unavailable"
+CATEGORY_INSUFFICIENT_MEMORY = "insufficient-memory"
+CATEGORY_INVALID_SHARD = "invalid-shard"
+CATEGORY_UNSUPPORTED_RECIPE = "unsupported-recipe"
+CATEGORY_LOAD_FAILED = "load-failed"
+CATEGORY_FORWARD_FAILED = "forward-failed"
+
+CATEGORY_HINTS: Mapping[str, str] = {
+    CATEGORY_NO_MODEL: (
+        "No model is selected. Pass --model <repo-or-name>, or run `meshnet-node` "
+        "once to save a config."
+    ),
+    CATEGORY_MISSING_DEPENDENCY: (
+        "The model runtime is not installed. Install the node's model extras "
+        "(torch, transformers, safetensors, accelerate, bitsandbytes)."
+    ),
+    CATEGORY_MODEL_UNAVAILABLE: (
+        "The model files could not be read. Check the model id, --download-dir, "
+        "and that the artifact is downloaded or reachable."
+    ),
+    CATEGORY_INSUFFICIENT_MEMORY: (
+        "This shard does not fit in memory. Serve fewer layers (--shard-start / "
+        "--shard-end) or use a smaller quantization (-q int8, -q nf4)."
+    ),
+    CATEGORY_INVALID_SHARD: (
+        "The requested layer range does not exist in this model. Check "
+        "--shard-start / --shard-end against the model's layer count."
+    ),
+    CATEGORY_UNSUPPORTED_RECIPE: (
+        "The recipe asks for an execution setting this backend cannot apply. "
+        "Select a different recipe with --recipe."
+    ),
+    CATEGORY_LOAD_FAILED: (
+        "The shard could not be loaded. Re-run with --debug for the full traceback."
+    ),
+    CATEGORY_FORWARD_FAILED: (
+        "The shard loaded but could not execute a forward pass. This node cannot "
+        "serve this model/shard; re-run with --debug for the full traceback."
+    ),
+}
+
+
+class DoctorError(RuntimeError):
+    """A validation failure with an operator-facing category and hint."""
+
+    def __init__(self, category: str, message: str) -> None:
+        super().__init__(message)
+        self.category = category
+
+    @property
+    def hint(self) -> str:
+        return CATEGORY_HINTS.get(self.category, "")
+
+
+@dataclass(frozen=True)
+class DoctorSelection:
+    """The one model/shard/config combination startup would load."""
+
+    model_id: str
+    shard_start: int
+    shard_end: int
+    quantization: str = "auto"
+    cache_dir: Path | None = None
+    force_cpu: bool = False
+
+    @property
+    def shard_label(self) -> str:
+        return f"layers {self.shard_start}–{self.shard_end}"
+
+
+@dataclass(frozen=True)
+class RecipeResult:
+    """One recipe's validation outcome, with the report it produced."""
+
+    recipe: Recipe
+    report: CapabilityReport
+    category: str | None = None
+    error: BaseException | None = None
+
+    @property
+    def passed(self) -> bool:
+        return self.report.passed
+
+    @property
+    def hint(self) -> str:
+        return CATEGORY_HINTS.get(self.category or "", "")
+
+
+@dataclass(frozen=True)
+class DoctorResult:
+    """The outcome of a doctor run over one or more recipes."""
+
+    selection: DoctorSelection
+    results: tuple[RecipeResult, ...] = ()
+
+    @property
+    def passed(self) -> bool:
+        return bool(self.results) and all(r.passed for r in self.results)
+
+    @property
+    def reports(self) -> tuple[CapabilityReport, ...]:
+        return tuple(r.report for r in self.results)
+
+    @property
+    def exit_code(self) -> int:
+        return 0 if self.passed else 1
+
+
+# --- selection: the same resolution startup performs ------------------------
+
+
+def resolve_selection(
+    cfg: Mapping[str, Any],
+    *,
+    detect_layers: Callable[[str, Path | None], int | None] | None = None,
+) -> DoctorSelection:
+    """Resolve config + flags into the selection startup would load.
+
+    This mirrors `startup.run_startup`: the same model id, the same
+    `bf16`→`bfloat16` quantization normalization, and the same shard default of
+    the whole model when no range is pinned. It deliberately does *not* ask the
+    tracker for a gap assignment — the doctor is an offline check of what this
+    node can run, and startup re-validates whatever range it is finally given.
+    """
+    model_id = _selected_model_id(cfg)
+    if not model_id:
+        raise DoctorError(
+            CATEGORY_NO_MODEL, "no model is selected in config or flags"
+        )
+
+    cache_dir = Path(cfg["download_dir"]) if cfg.get("download_dir") else None
+    quantization = str(cfg.get("quantization") or "auto").replace("bf16", "bfloat16")
+
+    shard_start = cfg.get("shard_start")
+    shard_end = cfg.get("shard_end")
+    if shard_start is None or shard_end is None:
+        detect = detect_layers or _detect_layers
+        total = detect(model_id, cache_dir)
+        if total is None:
+            raise DoctorError(
+                CATEGORY_MODEL_UNAVAILABLE,
+                f"could not read the layer count from the {model_id} config; "
+                "pass --shard-start and --shard-end explicitly",
+            )
+        shard_start = 0 if shard_start is None else shard_start
+        shard_end = total - 1 if shard_end is None else shard_end
+
+    if shard_start < 0 or shard_end < shard_start:
+        raise DoctorError(
+            CATEGORY_INVALID_SHARD,
+            f"invalid shard range {shard_start}–{shard_end}: start must be "
+            "non-negative and not greater than end",
+        )
+
+    return DoctorSelection(
+        model_id=model_id,
+        shard_start=int(shard_start),
+        shard_end=int(shard_end),
+        quantization=quantization,
+        cache_dir=cache_dir,
+        force_cpu=bool(cfg.get("force_cpu", False)),
+    )
+
+
+def _selected_model_id(cfg: Mapping[str, Any]) -> str | None:
+    """The HF repo startup would load, resolving a catalog alias if needed."""
+    hf_repo = str(cfg.get("model_hf_repo") or "").strip()
+    if hf_repo:
+        return hf_repo
+    name = str(cfg.get("model_name") or "").strip()
+    if not name:
+        return None
+    from .model_catalog import resolve_model_alias
+
+    preset = resolve_model_alias(name)
+    if preset is not None and preset.hf_repo:
+        return preset.hf_repo
+    return name if "/" in name else None
+
+
+def _detect_layers(model_id: str, cache_dir: Path | None) -> int | None:
+    from .startup import _detect_num_layers
+
+    return _detect_num_layers(model_id, cache_dir=cache_dir)
+
+
+# --- the bounded real forward ----------------------------------------------
+
+
+@dataclass(frozen=True)
+class ProbeInput:
+    """A synthetic hidden-state payload in the same wire format peers send."""
+
+    body: bytes
+    shape: list[int]
+    attention_mask_header: str | None
+    position_ids_header: str | None
+
+
+def _int64_header(rows: Sequence[Sequence[int]]) -> str:
+    """Encode an int64 tensor as `shape:base64`, matching the backend's format."""
+    flat = [int(v) for row in rows for v in row]
+    raw = struct.pack(f"<{len(flat)}q", *flat)
+    shape = f"{len(rows)},{len(rows[0])}" if rows else "0"
+    return f"{shape}:{base64.b64encode(raw).decode('ascii')}"
+
+
+def build_probe_input(hidden_size: int, tokens: int = PROBE_TOKENS) -> ProbeInput:
+    """Build a bounded mid-shard probe: `tokens` positions of bfloat16 zeros.
+
+    Zeros are a legitimate hidden state; what is being proven is that the
+    layers execute on this device, not that the output means anything. The
+    payload is built with plain bytes so callers need no Torch import.
+    """
+    if hidden_size <= 0:
+        raise DoctorError(
+            CATEGORY_FORWARD_FAILED,
+            "the backend reports no hidden size, so no probe tensor can be built",
+        )
+    ones = [[1] * tokens]
+    positions = [list(range(tokens))]
+    return ProbeInput(
+        body=b"\x00" * (tokens * hidden_size * 2),  # bfloat16 == 2 bytes
+        shape=[1, tokens, hidden_size],
+        attention_mask_header=_int64_header(ones),
+        position_ids_header=_int64_header(positions),
+    )
+
+
+def probe_forward(backend: Any, *, tokens: int = PROBE_TOKENS) -> dict:
+    """Run one bounded real forward through the shard `backend` holds.
+
+    Returns a small detail dict for the human summary. Raises `DoctorError`
+    (category `forward-failed`) if the shard cannot execute or returns nothing.
+    """
+    is_head = bool(getattr(backend, "is_head", False))
+    is_tail = bool(getattr(backend, "is_tail", False))
+
+    try:
+        if is_head:
+            output = backend.encode_prompt(PROBE_PROMPT)
+            kind = "prompt"
+            if is_tail:
+                # A head+tail shard owns the lm_head too. Re-entering above the
+                # last layer runs no layer again — it only decodes — so the whole
+                # selected shard is covered without a second forward through it.
+                output = backend.forward_bytes(
+                    output.body,
+                    output.shape,
+                    output.attention_mask_header,
+                    output.position_ids_header,
+                    start_layer=int(getattr(backend, "shard_end", 0)) + 1,
+                )
+                kind = "prompt+decode"
+        else:
+            probe = build_probe_input(int(getattr(backend, "hidden_size", 0) or 0))
+            output = backend.forward_bytes(
+                probe.body,
+                probe.shape,
+                probe.attention_mask_header,
+                probe.position_ids_header,
+                start_layer=getattr(backend, "shard_start", None),
+            )
+            kind = "hidden-states"
+    except DoctorError:
+        raise
+    except Exception as exc:
+        raise DoctorError(CATEGORY_FORWARD_FAILED, _describe(exc)) from exc
+
+    return {"probe": kind, "tokens": tokens, **_describe_output(output)}
+
+
+def _describe_output(output: Any) -> dict:
+    """Validate the forward produced real output, and summarize it."""
+    if output is None:
+        raise DoctorError(
+            CATEGORY_FORWARD_FAILED, "the shard forward returned no output"
+        )
+
+    token_id = getattr(output, "token_id", None)
+    if token_id is not None:  # tail shard: decoded a token
+        return {"output": "token", "token_id": int(token_id)}
+
+    body = getattr(output, "body", None)
+    shape = list(getattr(output, "shape", []) or [])
+    if not body or not shape:
+        raise DoctorError(
+            CATEGORY_FORWARD_FAILED,
+            "the shard forward returned an empty hidden-state payload",
+        )
+    return {"output": "hidden-states", "shape": shape}
+
+
+# --- running the doctor -----------------------------------------------------
+
+
+def default_load_backend(
+    selection: DoctorSelection,
+    recipe: Recipe,
+) -> Any:
+    """Load the shard through the exact path startup uses."""
+    from .torch_server import _load_backend
+
+    return _load_backend(
+        selection.model_id,
+        selection.shard_start,
+        selection.shard_end,
+        selection.quantization,
+        selection.cache_dir,
+        force_cpu=selection.force_cpu,
+        recipe_params=recipe.params,
+    )
+
+
+def select_recipes(
+    manifest: RecipeManifest,
+    *,
+    recipe_id: str | None = None,
+    all_recipes: bool = False,
+) -> tuple[Recipe, ...]:
+    """The recipes to validate: the selected one, or every one on request.
+
+    `--all-recipes` is the only way to pay for validating recipes the node was
+    not asked to serve; ordinary onboarding validates exactly one.
+    """
+    if all_recipes:
+        if recipe_id is not None:
+            raise DoctorError(
+                CATEGORY_UNSUPPORTED_RECIPE,
+                "--recipe and --all-recipes are mutually exclusive",
+            )
+        return manifest.recipes
+    try:
+        return (manifest.require(recipe_id or DEFAULT_RECIPE_ID),)
+    except RecipeManifestError as exc:
+        raise DoctorError(CATEGORY_UNSUPPORTED_RECIPE, str(exc)) from exc
+
+
+def run_doctor(
+    selection: DoctorSelection,
+    *,
+    manifest: RecipeManifest | None = None,
+    recipe_id: str | None = None,
+    all_recipes: bool = False,
+    load_backend: Callable[[DoctorSelection, Recipe], Any] | None = None,
+    now: Callable[[], float] | None = None,
+) -> DoctorResult:
+    """Validate the selection, one bounded real forward per recipe.
+
+    Never raises for a validation failure: every recipe yields a report, passed
+    or failed, so the caller can write the evidence out either way. `DoctorError`
+    only escapes for input the caller got wrong (an unknown recipe id).
+    """
+    manifest = manifest or load_recipe_manifest()
+    recipes = select_recipes(manifest, recipe_id=recipe_id, all_recipes=all_recipes)
+    clock = now or time.time
+    load = load_backend or default_load_backend
+
+    results = [
+        _validate_recipe(selection, recipe, manifest, load, clock)
+        for recipe in recipes
+    ]
+    return DoctorResult(selection=selection, results=tuple(results))
+
+
+def validate_loaded_backend(
+    backend: Any,
+    selection: DoctorSelection,
+    recipe: Recipe,
+    manifest: RecipeManifest,
+    *,
+    now: Callable[[], float] | None = None,
+) -> RecipeResult:
+    """Validate a shard that is already loaded, without loading it a second time.
+
+    Startup calls this on the very backend that would serve traffic, so the proof
+    it produces is about that object, not about a re-load that might have landed
+    on a different device.
+    """
+    return _validate_recipe(
+        selection, recipe, manifest, lambda *_: backend, now or time.time
+    )
+
+
+def _validate_recipe(
+    selection: DoctorSelection,
+    recipe: Recipe,
+    manifest: RecipeManifest,
+    load_backend: Callable[[DoctorSelection, Recipe], Any],
+    clock: Callable[[], float],
+) -> RecipeResult:
+    started = time.monotonic()
+    backend: Any = None
+    category: str | None = None
+    error: BaseException | None = None
+    diagnostics: list[str] = []
+    detail: dict = {}
+
+    try:
+        backend = load_backend(selection, recipe)
+        detail = probe_forward(backend)
+    except DoctorError as exc:
+        category, error = exc.category, exc
+        diagnostics = [str(exc), exc.hint]
+    except Exception as exc:  # noqa: BLE001 — every failure becomes a report
+        category = classify_failure(exc)
+        error = exc
+        diagnostics = [_describe(exc), CATEGORY_HINTS.get(category, "")]
+    duration_ms = int((time.monotonic() - started) * 1000)
+
+    device = _backend_device(backend, selection)
+    report = build_capability_report(
+        model_id=selection.model_id,
+        shard_start=selection.shard_start,
+        shard_end=selection.shard_end,
+        recipe_id=recipe.id,
+        recipe_version=recipe.version,
+        catalogue_version=manifest.catalogue_version,
+        backend_id=recipe.backend_id,
+        device=device,
+        device_name=_backend_device_name(device),
+        quantization=selection.quantization,
+        runtime=_runtime_versions(),
+        model_config=_model_config(backend),
+        status=STATUS_FAILED if category else STATUS_PASSED,
+        duration_ms=duration_ms,
+        diagnostics=[d for d in diagnostics if d] or None,
+        validated_at=clock(),
+    )
+    if category:
+        return RecipeResult(
+            recipe=recipe, report=report, category=category, error=error
+        )
+    return RecipeResult(recipe=recipe, report=report)
+
+
+def classify_failure(exc: BaseException) -> str:
+    """Map a backend exception to an operator-facing category.
+
+    Matches on the backend's own error types, never on model or vendor names.
+    """
+    from .model_backend import (
+        InsufficientVRAMError,
+        MissingModelDependencyError,
+        PartialModelLoadUnsupported,
+        UnsupportedRecipeParam,
+    )
+
+    if isinstance(exc, MissingModelDependencyError):
+        return CATEGORY_MISSING_DEPENDENCY
+    if isinstance(exc, InsufficientVRAMError):
+        return CATEGORY_INSUFFICIENT_MEMORY
+    if isinstance(exc, UnsupportedRecipeParam):
+        return CATEGORY_UNSUPPORTED_RECIPE
+    if isinstance(exc, PartialModelLoadUnsupported):
+        return CATEGORY_LOAD_FAILED
+    if isinstance(exc, ValueError):  # shard range vs. the model's real layers
+        return CATEGORY_INVALID_SHARD
+    if isinstance(exc, (FileNotFoundError, OSError)):
+        return CATEGORY_MODEL_UNAVAILABLE
+    return CATEGORY_LOAD_FAILED
+
+
+def _describe(exc: BaseException) -> str:
+    """A one-line, traceback-free description. Sanitized by the report."""
+    text = str(exc).strip()
+    return f"{type(exc).__name__}: {text}" if text else type(exc).__name__
+
+
+def _backend_device(backend: Any, selection: DoctorSelection) -> str:
+    device = getattr(backend, "device", None)
+    if device is None:
+        # The load failed, so no device was chosen — record the one that was asked for.
+        return "cpu" if selection.force_cpu else "unknown"
+    return str(getattr(device, "type", device))
+
+
+def _backend_device_name(device: str) -> str | None:
+    """The accelerator's name, when the shard actually landed on one."""
+    if device != "cuda":
+        return None
+    from .hardware import detect_hardware
+
+    try:
+        return detect_hardware().get("gpu_name") or None
+    except Exception:
+        return None
+
+
+def _model_config(backend: Any) -> Any:
+    """The loaded model's config, for the report's fingerprint."""
+    config = getattr(getattr(backend, "model", None), "config", None)
+    to_dict = getattr(config, "to_dict", None)
+    if not callable(to_dict):
+        return None
+    try:
+        return to_dict()
+    except Exception:
+        return None
+
+
+def _runtime_versions() -> dict[str, str]:
+    """Versions of the stack that ran the forward — opaque labels, never branches."""
+    versions: dict[str, str] = {}
+    for name in ("torch", "transformers"):
+        try:
+            module = __import__(name)
+        except Exception:
+            continue
+        version = getattr(module, "__version__", None)
+        if version:
+            versions[name] = str(version)
+    return versions
+
+
+# --- output -----------------------------------------------------------------
+
+DEFAULT_REPORT_FILENAME = "capability.json"
+
+
+def default_report_path() -> Path:
+    from .config import config_path
+
+    return config_path().parent / DEFAULT_REPORT_FILENAME
+
+
+def write_reports(reports: Sequence[CapabilityReport], path: Path) -> Path:
+    """Write the capability report(s) as JSON. A failed run writes too."""
+    import json
+
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if len(reports) == 1:
+        path.write_text(reports[0].to_json(indent=2) + "\n", encoding="utf-8")
+    else:
+        payload = [r.to_dict() for r in reports]
+        path.write_text(
+            json.dumps(payload, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+        )
+    return path
+
+
+def render_result(result: DoctorResult, *, report_path: Path | None = None) -> str:
+    """The human summary: what was validated, what to do if it failed."""
+    selection = result.selection
+    lines = [
+        "meshnet-node doctor",
+        f"  Model:        {selection.model_id}",
+        f"  Shard:        {selection.shard_label}",
+        f"  Quantization: {selection.quantization}",
+        "",
+    ]
+
+    for item in result.results:
+        mark = "PASS" if item.passed else "FAIL"
+        device = item.report.backend.device
+        lines.append(
+            f"  [{mark}] recipe {item.recipe.id} (v{item.recipe.version}) "
+            f"on {device} — {item.report.duration_ms} ms"
+        )
+        if not item.passed:
+            for diagnostic in item.report.diagnostics:
+                lines.append(f"         {diagnostic}")
+
+    lines.append("")
+    if result.passed:
+        count = len(result.results)
+        what = "recipe" if count == 1 else "recipes"
+        lines.append(
+            f"  OK — the selected shard ran a real forward for {count} {what}."
+        )
+    else:
+        failed = [r for r in result.results if not r.passed]
+        categories = ", ".join(dict.fromkeys(r.category or "unknown" for r in failed))
+        lines.append(f"  FAILED — {categories}. This node cannot serve this shard.")
+
+    if report_path is not None:
+        lines.append(f"  Capability report: {report_path}")
+    return "\n".join(lines)

packages/node/meshnet_node/model_backend.py

@@ -9,16 +9,26 @@ import json
 import os
 import threading
 import time
+import warnings
 from pathlib import Path
-from typing import Any, Literal
+from typing import Any, Literal, Mapping
 
 Quantization = Literal["auto", "bfloat16", "int8", "nf4"]
 
+# Recipe params this backend knows how to apply (see meshnet_node.recipe_manifest).
+# A recipe is only meaningful if its params actually reach the execution path, so
+# an unknown key is an error rather than a silent no-op.
+SUPPORTED_RECIPE_PARAMS = ("attn_implementation", "use_cache")
+
 
 class ModelBackendError(RuntimeError):
     """Base class for real model backend startup and execution failures."""
 
 
+class UnsupportedRecipeParam(ModelBackendError):
+    """Raised when a recipe asks for an execution param this backend cannot apply."""
+
+
 class MissingModelDependencyError(ModelBackendError):
     """Raised when optional model dependencies are not installed."""
 
@@ -61,6 +71,14 @@ def _torch_cuda_is_executable(torch_module: Any) -> bool:
 
 @dataclass(frozen=True)
 class TensorPayload:
+    """An immutable, request-owned binary activation payload.
+
+    ``body`` is always the exact bfloat16 wire body.  It is intentionally
+    owned bytes rather than a view into a request buffer so a payload can move
+    across a hop without retaining an HTTP/WebSocket frame after that request
+    completes.
+    """
+
     body: bytes
     shape: list[int]
     attention_mask_header: str | None
@@ -213,6 +231,7 @@ class TorchModelShard:
         quantization: Quantization = "auto",
         cache_dir: Path | None = None,
         force_cpu: bool = False,
+        recipe_params: Mapping[str, Any] | None = None,
     ) -> None:
         if shard_start < 0 or shard_end < 0 or shard_start > shard_end:
             raise ValueError("shard_start must be <= shard_end and non-negative")
@@ -220,6 +239,8 @@ class TorchModelShard:
         self.shard_start = shard_start
         self.shard_end = shard_end
         self.quantization = quantization
+        self.recipe_params = validate_recipe_params(recipe_params)
+        attn_implementation = self.recipe_params.get("attn_implementation")
 
         try:
             import torch
@@ -260,6 +281,7 @@ class TorchModelShard:
                     shard_end,
                     dtype,
                     self.device,
+                    attn_implementation=attn_implementation,
                 )
             else:
                 load_kwargs = {
@@ -270,6 +292,8 @@ class TorchModelShard:
                 }
                 if quant_config is not None:
                     load_kwargs["quantization_config"] = quant_config
+                if attn_implementation is not None:
+                    load_kwargs["attn_implementation"] = attn_implementation
                 self.model = AutoModelForCausalLM.from_pretrained(
                     load_source,
                     **load_kwargs,
@@ -313,6 +337,8 @@ class TorchModelShard:
         # consume CPU tensors ("Pointer argument cannot be accessed from Triton"),
         # so CPU shards intentionally stay on the stateless prefill path.
         self.supports_kv_cache = self.device.type != "cpu"
+        if self.recipe_params.get("use_cache") is False:
+            self.supports_kv_cache = False
         self.kv_sessions = SessionCacheStore(
             max_sessions=int(os.environ.get("MESHNET_KV_MAX_SESSIONS", "8")),
             ttl_seconds=float(os.environ.get("MESHNET_KV_TTL_SECONDS", "600")),
@@ -688,6 +714,19 @@ class TorchModelShard:
         )
 
 
+def validate_recipe_params(params: Mapping[str, Any] | None) -> dict[str, Any]:
+    """Return recipe params this backend can honour, or raise naming the bad key."""
+    if not params:
+        return {}
+    unsupported = [key for key in params if key not in SUPPORTED_RECIPE_PARAMS]
+    if unsupported:
+        raise UnsupportedRecipeParam(
+            f"recipe param(s) {', '.join(sorted(unsupported))} are not supported by this "
+            f"backend; it applies: {', '.join(SUPPORTED_RECIPE_PARAMS)}"
+        )
+    return dict(params)
+
+
 def load_torch_shard(
     model_id: str,
     shard_start: int,
@@ -695,9 +734,16 @@ def load_torch_shard(
     quantization: Quantization = "auto",
     cache_dir: Path | None = None,
     force_cpu: bool = False,
+    recipe_params: Mapping[str, Any] | None = None,
 ) -> TorchModelShard:
     return TorchModelShard(
-        model_id, shard_start, shard_end, quantization, cache_dir, force_cpu=force_cpu
+        model_id,
+        shard_start,
+        shard_end,
+        quantization,
+        cache_dir,
+        force_cpu=force_cpu,
+        recipe_params=recipe_params,
     )
 
 
@@ -747,6 +793,7 @@ def _load_partial_model_from_snapshot(
     init_empty_weights_fn: Any | None = None,
     set_tensor_fn: Any | None = None,
     safe_open_fn: Any | None = None,
+    attn_implementation: str | None = None,
 ) -> Any:
     from .model_catalog import layers_from_config
     from .safetensors_selection import (
@@ -763,6 +810,10 @@ def _load_partial_model_from_snapshot(
 
     snapshot_dir = Path(load_source)
     cfg = auto_config.from_pretrained(str(snapshot_dir))
+    if attn_implementation is not None:
+        # The partial path instantiates from the config, so the attention choice
+        # has to be set on it rather than passed to from_pretrained.
+        cfg._attn_implementation = attn_implementation
     total_layers = layers_from_config(cfg)
     if total_layers is None:
         raise PartialModelLoadUnsupported(
@@ -1120,7 +1171,21 @@ def _tensor_to_bytes(tensor: Any) -> bytes:
 
 
 def _tensor_from_bfloat16_bytes(body: bytes, shape: list[int], torch: Any) -> Any:
-    tensor = torch.frombuffer(bytearray(body), dtype=torch.bfloat16)
+    # ``frombuffer`` views the immutable request-owned bytes for this forward
+    # only.  The following device transfer is the one required CPU→GPU copy;
+    # wrapping in ``bytearray`` first used to add an avoidable CPU allocation
+    # and copy.  Do not upcast through float32: the activation wire contract
+    # is bfloat16 and model layers accept it directly.
+    # PyTorch warns because bytes are immutable even though the forward path
+    # never mutates this view.  Suppress only that known warning; copying into
+    # a writable bytearray would defeat the zero-copy decode path.
+    with warnings.catch_warnings():
+        warnings.filterwarnings(
+            "ignore",
+            message="The given buffer is not writable.*",
+            category=UserWarning,
+        )
+        tensor = torch.frombuffer(body, dtype=torch.bfloat16)
     return tensor.reshape(shape)
 
 

packages/node/meshnet_node/relay_bridge.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import base64
+import http.client
 import json
 import logging
 import os
@@ -10,8 +11,6 @@ import re
 import threading
 import time
 import urllib.parse
-import urllib.error
-import urllib.request
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
 
@@ -44,8 +43,14 @@ BINARY_FRAME_MAGIC = b"MRF1"
 
 
 def encode_binary_frame(header: dict, body: bytes) -> bytes:
+    """Build one request-owned binary frame without base64 expansion.
+
+    ``join`` makes one owned output frame rather than creating intermediate
+    concatenation frames.  The layout is intentionally unchanged because the
+    relay ships an independent copy of this codec.
+    """
     header_bytes = json.dumps(header, separators=(",", ":")).encode()
-    return BINARY_FRAME_MAGIC + len(header_bytes).to_bytes(4, "big") + header_bytes + body
+    return b"".join((BINARY_FRAME_MAGIC, len(header_bytes).to_bytes(4, "big"), header_bytes, body))
 
 
 def decode_binary_frame(frame: bytes) -> tuple[dict, bytes]:
@@ -53,7 +58,9 @@ def decode_binary_frame(frame: bytes) -> tuple[dict, bytes]:
         raise ValueError("not a meshnet binary relay frame")
     header_len = int.from_bytes(frame[4:8], "big")
     header = json.loads(frame[8:8 + header_len].decode())
-    return header, bytes(frame[8 + header_len:])
+    # The slice is a request-owned body.  It cannot retain the enclosing relay
+    # frame after callers finish processing it.
+    return header, frame[8 + header_len:]
 
 
 @dataclass(frozen=True)
@@ -82,6 +89,62 @@ def _max_concurrency_from_env() -> int:
     return max(1, value)
 
 
+class _LoopbackHttpClientPool:
+    """Bounded worker-local HTTP/1.1 clients for relay loopback forwarding."""
+
+    def __init__(self, base_url: str, timeout: float = 300.0) -> None:
+        parsed = urllib.parse.urlsplit(base_url.rstrip("/"))
+        if parsed.scheme not in {"http", "https"} or not parsed.hostname:
+            raise ValueError(f"invalid local bridge URL: {base_url!r}")
+        self._scheme = parsed.scheme
+        self._host = parsed.hostname
+        self._port = parsed.port
+        self._base_path = parsed.path.rstrip("/")
+        self._timeout = timeout
+        self._local = threading.local()
+        self._lock = threading.Lock()
+        self._clients: set[http.client.HTTPConnection] = set()
+
+    def _connection(self) -> http.client.HTTPConnection:
+        connection = getattr(self._local, "connection", None)
+        if connection is None:
+            kind = http.client.HTTPSConnection if self._scheme == "https" else http.client.HTTPConnection
+            connection = kind(self._host, self._port, timeout=self._timeout)
+            self._local.connection = connection
+            with self._lock:
+                self._clients.add(connection)
+        return connection
+
+    def request(self, method: str, path: str, body: bytes, headers: dict):
+        request_path = f"{self._base_path}{path if path.startswith('/') else '/' + path}"
+        connection = self._connection()
+        try:
+            connection.request(method, request_path, body=body, headers=headers)
+            return connection.getresponse()
+        except Exception:
+            self.discard()
+            raise
+
+    def discard(self) -> None:
+        connection = getattr(self._local, "connection", None)
+        if connection is None:
+            return
+        try:
+            connection.close()
+        finally:
+            self._local.connection = None
+            with self._lock:
+                self._clients.discard(connection)
+
+    def close(self) -> None:
+        with self._lock:
+            clients = tuple(self._clients)
+            self._clients.clear()
+        for connection in clients:
+            connection.close()
+        self._local.connection = None
+
+
 class RelayHttpBridge:
     """Connect outbound to a relay and proxy relay HTTP requests to localhost.
 
@@ -115,6 +178,7 @@ class RelayHttpBridge:
         self._decode_log_lock = threading.Lock()
         self._decode_steps: dict[str, int] = {}
         self._ws = None
+        self._loopback_clients = _LoopbackHttpClientPool(self.local_base_url)
 
     @property
     def relay_addr(self) -> str:
@@ -141,6 +205,7 @@ class RelayHttpBridge:
             self._thread.join(timeout=3.0)
         if self._executor is not None:
             self._executor.shutdown(wait=False)
+        self._loopback_clients.close()
 
     def _run(self) -> None:
         import websockets.sync.client as wsc  # type: ignore[import]
@@ -260,14 +325,14 @@ class RelayHttpBridge:
                 body_text = payload.get("body") or ""
                 data = body_text.encode() if isinstance(body_text, str) else bytes(body_text)
 
-        url = f"{self.local_base_url}{path}"
-        req = urllib.request.Request(url, data=data, headers=headers, method=method)
         try:
-            with urllib.request.urlopen(req, timeout=300.0) as resp:
+            resp = self._loopback_clients.request(method, path, data, headers)
+            try:
                 resp_headers = dict(resp.headers)
                 content_type = resp.headers.get("Content-Type", "")
                 if "text/event-stream" in content_type:
-                    self._stream_response(request_id, resp, resp_headers)
+                    if not self._stream_response(request_id, resp, resp_headers):
+                        self._loopback_clients.discard()
                     return
                 resp_bytes = resp.read()
                 # Forward all X-Meshnet-* headers so the caller can reconstruct the activation.
@@ -289,14 +354,18 @@ class RelayHttpBridge:
                 else:
                     result["body"] = resp_bytes.decode(errors="replace")
                 self._send_response_frame(result)
-        except urllib.error.HTTPError as exc:
+            finally:
+                resp.close()
+        except http.client.HTTPException as exc:
+            self._loopback_clients.discard()
             self._send_response_frame({
                 "request_id": request_id,
-                "status": exc.code,
-                "headers": {"Content-Type": exc.headers.get("Content-Type", "application/json")},
-                "body": exc.read().decode(errors="replace"),
+                "status": 503,
+                "headers": {"Content-Type": "application/json"},
+                "body": json.dumps({"error": f"relay bridge local request failed: {exc}"}),
             })
         except Exception as exc:
+            self._loopback_clients.discard()
             self._send_response_frame({
                 "request_id": request_id,
                 "status": 503,
@@ -304,7 +373,7 @@ class RelayHttpBridge:
                 "body": json.dumps({"error": f"relay bridge local request failed: {exc}"}),
             })
 
-    def _stream_response(self, request_id: str, resp, resp_headers: dict) -> None:
+    def _stream_response(self, request_id: str, resp, resp_headers: dict) -> bool:
         """Forward an SSE response as chunk frames, one per complete SSE event.
 
         Frame order: header frame (status + headers), chunk frames, done frame.
@@ -319,7 +388,7 @@ class RelayHttpBridge:
             "done": False,
         })
         if not sent:
-            return
+            return False
         event_lines: list[str] = []
         for raw_line in resp:
             line = raw_line.decode(errors="replace")
@@ -333,7 +402,7 @@ class RelayHttpBridge:
                 "chunk": "".join(event_lines),
                 "done": False,
             }):
-                return
+                return False
             event_lines = []
         if event_lines:
             if not self._send_response_frame({
@@ -342,8 +411,8 @@ class RelayHttpBridge:
                 "chunk": "".join(event_lines),
                 "done": False,
             }):
-                return
-        self._send_response_frame({
+                return False
+        return self._send_response_frame({
             "request_id": request_id,
             "stream": True,
             "done": True,

packages/node/meshnet_node/route_session_benchmark.py

@@ -0,0 +1,385 @@
+"""Deterministic, stub-backed Route Session transport benchmark.
+
+This is deliberately a transport harness, not a model benchmark.  It gives
+performance work a repeatable baseline without requiring a GPU, a live relay,
+or localhost sockets (which are not available in every CI sandbox).
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+import urllib.request
+import zlib
+from collections import defaultdict
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Iterable, Literal
+
+TransportMode = Literal["direct", "relay"]
+CacheMode = Literal["cached", "stateless"]
+
+
+@dataclass(frozen=True)
+class BenchmarkScenario:
+    """Fixed input and expected output for one reproducible Route Session."""
+
+    prompt: str = "Route Session profiling prompt."
+    output_tokens: tuple[str, ...] = (" amber", " birch", " cedar", " dogwood")
+    activation_bytes: int = 4096
+    compression: bool = True
+
+
+@dataclass(frozen=True)
+class SeamSample:
+    """One head-to-tail activation transfer, with all durations in milliseconds."""
+
+    phase: Literal["prefill", "decode"]
+    token_index: int | None
+    session_id: str
+    activation_id: str
+    seam: str
+    mode: TransportMode
+    cache_mode: CacheMode
+    model_ms: float
+    encode_ms: float
+    framing_ms: float
+    metadata_ms: float
+    copy_allocation_ms: float
+    copy_allocation_bytes: int
+    compression_ms: float
+    decompression_ms: float
+    connection_setup_ms: float
+    queue_wait_ms: float
+    transport_ms: float
+    seam_latency_ms: float
+    payload_bytes: int
+    wire_bytes: int
+    compression_ratio: float
+    connection_attempted: bool
+
+
+@dataclass(frozen=True)
+class BenchmarkRun:
+    """JSON-safe result for one mode/cache-mode scenario."""
+
+    scenario: BenchmarkScenario
+    mode: TransportMode
+    cache_mode: CacheMode
+    output_tokens: tuple[str, ...]
+    samples: tuple[SeamSample, ...]
+    cleanup: dict[str, int | bool]
+
+    def to_dict(self) -> dict:
+        samples = [asdict(sample) for sample in self.samples]
+        return {
+            "scenario": asdict(self.scenario),
+            "mode": self.mode,
+            "cache_mode": self.cache_mode,
+            "output_tokens": list(self.output_tokens),
+            "session_id": self.samples[0].session_id if self.samples else "",
+            "cleanup": self.cleanup,
+            "connections": {
+                "attempts": sum(sample.connection_attempted for sample in self.samples),
+            },
+            "phases": _summaries_by(self.samples, lambda sample: sample.phase),
+            "seams": _summaries_by(self.samples, lambda sample: sample.seam),
+            "samples": samples,
+        }
+
+
+def _percentile(values: Iterable[float], percentile: float) -> float:
+    ordered = sorted(values)
+    if not ordered:
+        return 0.0
+    index = max(0, (len(ordered) * percentile + 99) // 100 - 1)
+    return round(ordered[int(index)], 4)
+
+
+def _summary(samples: list[SeamSample]) -> dict[str, float | int]:
+    total_latency_ms = sum(sample.seam_latency_ms for sample in samples)
+    return {
+        "count": len(samples),
+        "p50_latency_ms": _percentile((sample.seam_latency_ms for sample in samples), 50),
+        "p95_latency_ms": _percentile((sample.seam_latency_ms for sample in samples), 95),
+        "payload_bytes": sum(sample.payload_bytes for sample in samples),
+        "wire_bytes": sum(sample.wire_bytes for sample in samples),
+        "compression_ratio": round(
+            sum(sample.payload_bytes for sample in samples) / max(1, sum(sample.wire_bytes for sample in samples)), 4
+        ),
+        "connection_attempts": sum(sample.connection_attempted for sample in samples),
+        "p50_queue_wait_ms": _percentile((sample.queue_wait_ms for sample in samples), 50),
+        "p95_queue_wait_ms": _percentile((sample.queue_wait_ms for sample in samples), 95),
+        "tokens_per_sec": round(
+            sum(sample.phase == "decode" for sample in samples) / max(0.001, total_latency_ms / 1000), 4
+        ),
+        "bytes_per_token": round(
+            sum(sample.wire_bytes for sample in samples) / max(1, sum(sample.phase == "decode" for sample in samples)), 4
+        ),
+        "compression_cpu_ms": round(
+            sum(sample.compression_ms + sample.decompression_ms for sample in samples), 4
+        ),
+        "peak_buffered_bytes": max((sample.copy_allocation_bytes for sample in samples), default=0),
+    }
+
+
+def _summaries_by(samples: tuple[SeamSample, ...], key) -> dict[str, dict[str, float | int]]:
+    groups: dict[str, list[SeamSample]] = defaultdict(list)
+    for sample in samples:
+        groups[key(sample)].append(sample)
+    return {name: _summary(group) for name, group in groups.items()}
+
+
+class _StubTransport:
+    """A deterministic two-node seam with explicit connection ownership."""
+
+    def __init__(self, mode: TransportMode, cache_mode: CacheMode, scenario: BenchmarkScenario) -> None:
+        self.mode = mode
+        self.cache_mode = cache_mode
+        self.scenario = scenario
+        self._open_connections: set[str] = set()
+        self.session_id = "benchmark-route-session"
+        self._activation_count = 0
+        self._closed = False
+
+    def transfer(self, phase: Literal["prefill", "decode"], token_index: int | None) -> SeamSample:
+        # Cached Route Sessions own one connection per seam in both direct and
+        # relay modes. Stateless calls deliberately remain one-shot baselines.
+        persistent = self.cache_mode == "cached"
+        request_key = "route-session" if persistent else f"{phase}:{token_index}"
+        connection_attempted = request_key not in self._open_connections
+        self._open_connections.add(request_key)
+        self._activation_count += 1
+
+        payload = _activation(self.scenario.activation_bytes, phase, token_index)
+        wire = zlib.compress(payload, level=9) if self.scenario.compression else payload
+        payload_bytes, wire_bytes = len(payload), len(wire)
+        connection_setup_ms = (0.8 if self.mode == "direct" else 1.4) if connection_attempted else 0.0
+        queue_wait_ms = 0.0 if self.mode == "direct" else 0.18 + (0.05 if token_index is not None and token_index % 2 else 0.0)
+        model_ms = 1.6 if phase == "prefill" else 0.45
+        encode_ms = 0.16 if phase == "prefill" else 0.06
+        # Keep framing/metadata/copy costs explicit rather than hiding them in
+        # serialization or transport time.  The stub owns one binary frame and
+        # one response body per hop; no base64 body is modeled.
+        framing_ms = 0.035 if phase == "prefill" else 0.012
+        metadata_ms = 0.018 if phase == "prefill" else 0.008
+        copy_allocation_ms = 0.025 if self.scenario.compression else 0.012
+        copy_allocation_bytes = wire_bytes + payload_bytes
+        compression_ms = 0.09 if self.scenario.compression else 0.0
+        decompression_ms = 0.07 if self.scenario.compression else 0.0
+        transport_ms = (0.32 if self.mode == "direct" else 0.61) + wire_bytes / 100_000
+        seam_latency_ms = round(
+            model_ms + encode_ms + framing_ms + metadata_ms + copy_allocation_ms
+            + compression_ms + decompression_ms + connection_setup_ms + queue_wait_ms + transport_ms,
+            4,
+        )
+        return SeamSample(
+            phase=phase, token_index=token_index, session_id=self.session_id,
+            activation_id=f"benchmark-activation-{self._activation_count}", seam="head->tail", mode=self.mode,
+            cache_mode=self.cache_mode, model_ms=model_ms, encode_ms=encode_ms,
+            framing_ms=framing_ms, metadata_ms=metadata_ms,
+            copy_allocation_ms=copy_allocation_ms, copy_allocation_bytes=copy_allocation_bytes,
+            compression_ms=compression_ms, decompression_ms=decompression_ms,
+            connection_setup_ms=connection_setup_ms, queue_wait_ms=queue_wait_ms,
+            transport_ms=round(transport_ms, 4), seam_latency_ms=seam_latency_ms,
+            payload_bytes=payload_bytes, wire_bytes=wire_bytes,
+            compression_ratio=round(payload_bytes / wire_bytes, 4), connection_attempted=connection_attempted,
+        )
+
+    def close(self) -> dict[str, int | bool]:
+        """Close all deterministic owners and expose a CI-checkable snapshot."""
+        self._open_connections.clear()
+        self._closed = True
+        return {
+            "session_closed": True,
+            "open_connections": 0,
+            "queued_activations": 0,
+            "telemetry_aggregates": 0,
+        }
+
+
+def _activation(size: int, phase: str, token_index: int | None) -> bytes:
+    """Return a compressible but phase-distinguishable activation body."""
+    prefix = f"{phase}:{token_index if token_index is not None else 'prompt'}:".encode()
+    return (prefix * ((size // len(prefix)) + 1))[:size]
+
+
+def run_route_session_benchmark(
+    mode: TransportMode,
+    cache_mode: CacheMode,
+    scenario: BenchmarkScenario = BenchmarkScenario(),
+) -> BenchmarkRun:
+    """Run one fixed two-node prefill + decode Route Session scenario."""
+    transport = _StubTransport(mode, cache_mode, scenario)
+    try:
+        samples = [transport.transfer("prefill", None)]
+        samples.extend(transport.transfer("decode", index) for index in range(len(scenario.output_tokens)))
+    finally:
+        cleanup = transport.close()
+    return BenchmarkRun(scenario, mode, cache_mode, scenario.output_tokens, tuple(samples), cleanup)
+
+
+def run_benchmark_matrix(scenario: BenchmarkScenario = BenchmarkScenario()) -> dict:
+    """Run direct/relay and cached/stateless baselines suitable for CI artifacts."""
+    runs = [
+        run_route_session_benchmark(mode, cache_mode, scenario).to_dict()
+        for mode in ("direct", "relay")
+        for cache_mode in ("cached", "stateless")
+    ]
+    return {"schema_version": 1, "runs": runs}
+
+
+def assert_benchmark(
+    run: BenchmarkRun,
+    *,
+    expected_tokens: Iterable[str],
+    expected_connection_attempts: int,
+) -> None:
+    """Assertion seam for regression tests and future performance gates."""
+    assert tuple(expected_tokens) == run.output_tokens, "stub output tokens changed"
+    actual_attempts = sum(sample.connection_attempted for sample in run.samples)
+    assert actual_attempts == expected_connection_attempts, (
+        f"expected {expected_connection_attempts} connections, got {actual_attempts}"
+    )
+
+
+@dataclass(frozen=True)
+class PerformanceThresholds:
+    """Stable gate limits.
+
+    A cached decode must retain at least a 20% latency/throughput advantage and
+    cannot add more than 20% wire bytes per token.  Those deliberately broad
+    ratios tolerate ordinary LAN host variance, yet still catch loss of
+    connection reuse or a material transport/data-plane slowdown.  Exact
+    correctness, ownership, and cleanup invariants are enforced separately.
+    """
+
+    max_cached_p50_latency_ratio: float = 0.80
+    min_cached_throughput_ratio: float = 1.20
+    max_bytes_per_token_ratio: float = 1.20
+
+
+def assert_performance_gate(
+    report: dict,
+    *,
+    thresholds: PerformanceThresholds = PerformanceThresholds(),
+) -> None:
+    """Fail CI on a material transport regression, not ordinary host variation.
+
+    The stub's timing is deterministic, but ratios deliberately allow 20% when
+    the report is later compared with a LAN capture.  Connection ownership,
+    token identity, Route Session stability, and post-run cleanup are exact
+    invariants and must never be relaxed.
+    """
+    runs = {(run["mode"], run["cache_mode"]): run for run in report["runs"]}
+    expected = BenchmarkScenario().output_tokens
+    for key, run in runs.items():
+        assert tuple(run["output_tokens"]) == expected, f"{key}: output tokens changed"
+        samples = run["samples"]
+        assert len({sample["session_id"] for sample in samples}) == 1, f"{key}: Route Session changed"
+        assert len({sample["activation_id"] for sample in samples}) == len(samples), f"{key}: activation IDs reused"
+        assert run["cleanup"] == {
+            "session_closed": True, "open_connections": 0,
+            "queued_activations": 0, "telemetry_aggregates": 0,
+        }, f"{key}: resources leaked"
+        expected_connections = 1 if key[1] == "cached" else len(samples)
+        assert run["connections"]["attempts"] == expected_connections, f"{key}: connection regression"
+
+    for mode in ("direct", "relay"):
+        cached = runs[(mode, "cached")]["phases"]["decode"]
+        stateless = runs[(mode, "stateless")]["phases"]["decode"]
+        assert cached["p50_latency_ms"] <= stateless["p50_latency_ms"] * thresholds.max_cached_p50_latency_ratio, (
+            f"{mode}: cached p50 latency regressed"
+        )
+        assert cached["tokens_per_sec"] >= stateless["tokens_per_sec"] * thresholds.min_cached_throughput_ratio, (
+            f"{mode}: cached throughput regressed"
+        )
+        assert cached["bytes_per_token"] <= stateless["bytes_per_token"] * thresholds.max_bytes_per_token_ratio, (
+            f"{mode}: cached bytes/token regressed"
+        )
+
+
+def run_real_model_lan_benchmark(url: str, *, model: str, timeout: float = 120.0) -> dict:
+    """Opt-in client-side LAN capture using the same report schema as CI.
+
+    This intentionally makes exactly one OpenAI-compatible request.  It is a
+    live validation aid, not a CI input: remote seam CPU/buffer values are zero
+    until nodes expose them in a response, while bytes, latency, output and
+    connection ownership are measured at the LAN client boundary.
+    """
+    scenario = BenchmarkScenario()
+    body = json.dumps({
+        "model": model,
+        "messages": [{"role": "user", "content": scenario.prompt}],
+        "max_tokens": len(scenario.output_tokens), "temperature": 0,
+    }).encode()
+    request = urllib.request.Request(
+        f"{url.rstrip('/')}/v1/chat/completions", data=body,
+        headers={"Content-Type": "application/json", "X-Meshnet-Session": "lan-benchmark-session"}, method="POST",
+    )
+    started = time.monotonic()
+    with urllib.request.urlopen(request, timeout=timeout) as response:
+        response_body = response.read()
+        session_id = response.headers.get("X-Meshnet-Session", "lan-benchmark-session")
+    elapsed_ms = round((time.monotonic() - started) * 1000, 4)
+    payload = json.loads(response_body)
+    content = payload["choices"][0]["message"]["content"]
+    tokens = tuple(content.split())
+    sample = SeamSample(
+        phase="decode", token_index=0, session_id=session_id, activation_id="lan-activation-1",
+        seam="head->tail", mode="direct", cache_mode="cached", model_ms=0.0, encode_ms=0.0,
+        framing_ms=0.0, metadata_ms=0.0, copy_allocation_ms=0.0, copy_allocation_bytes=0,
+        compression_ms=0.0, decompression_ms=0.0, connection_setup_ms=elapsed_ms,
+        queue_wait_ms=0.0, transport_ms=elapsed_ms, seam_latency_ms=elapsed_ms,
+        payload_bytes=len(body), wire_bytes=len(body) + len(response_body), compression_ratio=1.0,
+        connection_attempted=True,
+    )
+    run = BenchmarkRun(
+        scenario, "direct", "cached", tokens, (sample,),
+        {"session_closed": True, "open_connections": 0, "queued_activations": 0, "telemetry_aggregates": 0},
+    )
+    return {"schema_version": 1, "source": "real-model-lan-client", "runs": [run.to_dict()]}
+
+
+def format_summary(report: dict) -> str:
+    """Render the compact, human-readable companion to the JSON artifact."""
+    lines = ["Route Session benchmark"]
+    for run in report["runs"]:
+        decode = run["phases"]["decode"]
+        seam = run["seams"]["head->tail"]
+        lines.append(
+            f"{run['mode']:6} {run['cache_mode']:9} "
+            f"decode p50/p95 {decode['p50_latency_ms']:.2f}/{decode['p95_latency_ms']:.2f} ms; "
+            f"{decode['tokens_per_sec']:.1f} tok/s; {decode['bytes_per_token']:.0f} B/tok; "
+            f"seam {seam['payload_bytes']}/{seam['wire_bytes']} B "
+            f"({seam['compression_ratio']:.2f}x); connections {run['connections']['attempts']}; "
+            f"queue p95 {decode['p95_queue_wait_ms']:.2f} ms"
+        )
+    return "\n".join(lines)
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Run the deterministic Route Session benchmark")
+    parser.add_argument("--json-out", type=Path, help="write the JSON artifact to this path")
+    parser.add_argument("--real-model-lan", metavar="URL", help="opt-in OpenAI-compatible LAN endpoint capture")
+    parser.add_argument("--model", help="model name required with --real-model-lan")
+    parser.add_argument("--timeout", type=float, default=120.0, help="LAN request timeout in seconds")
+    parser.add_argument("--no-gate", action="store_true", help="report deterministic results without enforcing thresholds")
+    args = parser.parse_args(argv)
+    if args.real_model_lan:
+        if not args.model:
+            parser.error("--model is required with --real-model-lan")
+        report = run_real_model_lan_benchmark(args.real_model_lan, model=args.model, timeout=args.timeout)
+    else:
+        report = run_benchmark_matrix()
+        if not args.no_gate:
+            assert_performance_gate(report)
+    if args.json_out:
+        args.json_out.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+    print(format_summary(report))
+    return 0
+
+
+if __name__ == "__main__":  # pragma: no cover - CLI entry point
+    raise SystemExit(main())

packages/node/meshnet_node/seam_telemetry.py

@@ -0,0 +1,155 @@
+"""Bounded, in-process telemetry for distributed activation seams.
+
+The generation path records one cheap counter update per activation.  It never
+flushes telemetry or performs I/O; callers decide when an aggregate should be
+logged or exposed to a heartbeat.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import time
+
+
+@dataclass
+class _SeamAggregate:
+    phase: str
+    hop: int
+    node: str
+    count: int = 0
+    latency_ms: float = 0.0
+    wire_bytes: int = 0
+    response_bytes: int = 0
+    compression_input_bytes: int = 0
+    compression_output_bytes: int = 0
+    compression_ms: float = 0.0
+    decompression_input_bytes: int = 0
+    decompression_output_bytes: int = 0
+    decompression_ms: float = 0.0
+    reused_connections: int = 0
+    last_activation_id: str = ""
+
+
+class GenerationTelemetry:
+    """Aggregate activation measurements for one stable Route Session."""
+
+    def __init__(
+        self,
+        session_id: str,
+        *,
+        report_every: int = 32,
+        report_interval: float = 5.0,
+        now: float | None = None,
+    ) -> None:
+        self.session_id = session_id
+        self.report_every = max(1, report_every)
+        self.report_interval = max(0.0, report_interval)
+        self.started = time.monotonic() if now is None else now
+        self._last_report = self.started
+        self._total_tokens = 0
+        self._closed = False
+        self._report_due = False
+        self._seams: dict[tuple[str, int, str], _SeamAggregate] = {}
+
+    def record_seam(
+        self,
+        *,
+        activation_id: str,
+        phase: str,
+        hop: int,
+        node: str,
+        latency_seconds: float,
+        wire_bytes: int,
+        response_bytes: int,
+        connection_reused: bool,
+        now: float | None = None,
+    ) -> bool:
+        """Record one activation locally and say whether a summary is due."""
+        if self._closed:
+            return False
+        observed = time.monotonic() if now is None else now
+        key = (phase, hop, node)
+        aggregate = self._seams.get(key)
+        if aggregate is None:
+            aggregate = _SeamAggregate(phase=phase, hop=hop, node=node)
+            self._seams[key] = aggregate
+        aggregate.count += 1
+        aggregate.latency_ms += max(0.0, latency_seconds) * 1000.0
+        aggregate.wire_bytes += max(0, wire_bytes)
+        aggregate.response_bytes += max(0, response_bytes)
+        aggregate.reused_connections += int(connection_reused)
+        aggregate.last_activation_id = activation_id
+        due = (
+            aggregate.count == 1
+            or aggregate.count % self.report_every == 0
+            or observed - self._last_report >= self.report_interval
+        )
+        self._report_due = self._report_due or due
+        return due
+
+    @property
+    def report_due(self) -> bool:
+        return self._report_due
+
+    def note_tokens(self, tokens: int) -> None:
+        if not self._closed:
+            self._total_tokens = max(0, tokens)
+
+    def record_compression(
+        self, *, phase: str, hop: int, node: str, input_bytes: int,
+        output_bytes: int, elapsed_seconds: float, decompression: bool = False,
+    ) -> None:
+        """Attach compression work to the same bounded seam aggregate."""
+        if self._closed:
+            return
+        key = (phase, hop, node)
+        aggregate = self._seams.get(key)
+        if aggregate is None:
+            aggregate = _SeamAggregate(phase=phase, hop=hop, node=node)
+            self._seams[key] = aggregate
+        if decompression:
+            aggregate.decompression_input_bytes += max(0, input_bytes)
+            aggregate.decompression_output_bytes += max(0, output_bytes)
+            aggregate.decompression_ms += max(0.0, elapsed_seconds) * 1000.0
+        else:
+            aggregate.compression_input_bytes += max(0, input_bytes)
+            aggregate.compression_output_bytes += max(0, output_bytes)
+            aggregate.compression_ms += max(0.0, elapsed_seconds) * 1000.0
+
+    def snapshot(self, *, now: float | None = None) -> dict:
+        observed = time.monotonic() if now is None else now
+        elapsed = max(observed - self.started, 1e-6)
+        seams = []
+        for aggregate in self._seams.values():
+            seams.append({
+                "phase": aggregate.phase,
+                "hop": aggregate.hop,
+                "node": aggregate.node,
+                "activations": aggregate.count,
+                "latency_ms": round(aggregate.latency_ms, 3),
+                "avg_latency_ms": round(aggregate.latency_ms / max(1, aggregate.count), 3),
+                "wire_bytes": aggregate.wire_bytes,
+                "response_bytes": aggregate.response_bytes,
+                "compression_input_bytes": aggregate.compression_input_bytes,
+                "compression_output_bytes": aggregate.compression_output_bytes,
+                "compression_ms": round(aggregate.compression_ms, 3),
+                "decompression_input_bytes": aggregate.decompression_input_bytes,
+                "decompression_output_bytes": aggregate.decompression_output_bytes,
+                "decompression_ms": round(aggregate.decompression_ms, 3),
+                "connection_reuse": aggregate.reused_connections,
+                "last_activation_id": aggregate.last_activation_id,
+            })
+        return {
+            "session_id": self.session_id,
+            "tokens_per_sec": round(self._total_tokens / elapsed, 2),
+            "seams": seams,
+        }
+
+    def mark_reported(self, *, now: float | None = None) -> None:
+        self._last_report = time.monotonic() if now is None else now
+        self._report_due = False
+
+    def close(self) -> None:
+        self._closed = True
+        self._seams.clear()
+        self._report_due = False

packages/node/meshnet_node/server.py

@@ -8,6 +8,7 @@ import urllib.parse
 from pathlib import Path
 
 from .downloader import compute_shard_checksum, write_shard_archive
+from .activation_compression import CompressionPolicies, compress_activation, decompress_activation
 
 # Binary activation wire format (contract for all shard nodes):
 # POST /forward with raw tensor bytes in the body and tensor/session/chunk
@@ -21,6 +22,7 @@ _DTYPE_SIZES = {
     "bfloat16": 2,
     "float32": 4,
 }
+_COMPRESSION_POLICIES = CompressionPolicies()
 
 
 def _make_stub_binary_activation(shape: list[int], dtype: str) -> bytes:
@@ -42,16 +44,7 @@ def _parse_shape(value: str | None) -> list[int]:
 
 
 def _decompress_body(body: bytes, encoding: str | None) -> bytes:
-    if not encoding:
-        return body
-    if encoding != "zstd":
-        raise ValueError("unsupported X-Meshnet-Encoding")
-    import zstandard as zstd
-
-    try:
-        return zstd.ZstdDecompressor().decompress(body)
-    except zstd.ZstdError as exc:
-        raise ValueError("invalid zstd activation body") from exc
+    return decompress_activation(body, encoding).body
 
 
 def _compress_body(body: bytes, encoding: str | None) -> bytes:
@@ -307,7 +300,14 @@ class _StubHandler(http.server.BaseHTTPRequestHandler):
             server.received_activations = True
 
         raw_payload = _make_stub_binary_activation(shape, dtype)
-        payload = _compress_body(raw_payload, encoding)
+        route_condition = self.headers.get("X-Meshnet-Compression-Route", "lan")
+        phase_condition = self.headers.get("X-Meshnet-Cache", "prefill")
+        if phase_condition not in {"prefill", "decode"}:
+            phase_condition = "prefill"
+        compression = compress_activation(
+            raw_payload, _COMPRESSION_POLICIES.for_condition(route_condition, phase_condition),
+        )
+        payload = compression.body
         self.send_response(200)
         self.send_header("Content-Type", "application/octet-stream")
         self.send_header("Content-Length", str(len(payload)))
@@ -317,8 +317,8 @@ class _StubHandler(http.server.BaseHTTPRequestHandler):
         self.send_header("X-Meshnet-Session", session)
         self.send_header("X-Meshnet-Chunk-Index", chunk_index)
         self.send_header("X-Meshnet-Chunk-Total", chunk_total)
-        if encoding:
-            self.send_header("X-Meshnet-Encoding", encoding)
+        if compression.encoding:
+            self.send_header("X-Meshnet-Encoding", compression.encoding)
         if server.is_last_shard:
             self.send_header("X-Meshnet-Stub-Response-Prefix", server.response_prefix)
         self.end_headers()

packages/node/meshnet_node/startup.py

@@ -14,9 +14,19 @@ import urllib.request
 from pathlib import Path
 from typing import Any
 
+from .admission import (
+    AdmissionRequirement,
+    CapabilityContext,
+    CapabilityValidator,
+    admit,
+    probe_capability,
+)
+from .capability import CapabilityReport
+from .doctor import DoctorSelection
 from .downloader import compute_shard_checksum, download_shard
 from .hardware import detect_hardware, benchmark_throughput_checked, with_forced_cpu
 from .model_catalog import model_metadata_for
+from .recipe_manifest import DEFAULT_RECIPE_ID, Recipe, RecipeManifest, load_recipe_manifest
 from .relay_bridge import RelayHttpBridge, peer_id_from_wallet
 from .server import StubNodeServer
 from .torch_server import TorchNodeServer
@@ -646,6 +656,68 @@ def _tracker_http_error_message(exc: urllib.error.HTTPError) -> str:
     return f"Tracker rejected shard assignment (HTTP {exc.code}): {detail}"
 
 
+def _resolve_recipe(recipe_id: str | None) -> tuple[RecipeManifest, Recipe]:
+    """The recipe this node will serve with — resolved before any weights load."""
+    manifest = load_recipe_manifest()
+    return manifest, manifest.require(recipe_id or DEFAULT_RECIPE_ID)
+
+
+def _capability_device(backend: Any, detected_device: str) -> str:
+    """The device the shard actually landed on, or the one this node detected."""
+    device = getattr(backend, "device", None)
+    if device is None:
+        return detected_device
+    return str(getattr(device, "type", device))
+
+
+def _admit_capability(
+    node: Any,
+    *,
+    model_id: str,
+    shard_start: int,
+    shard_end: int,
+    quantization: str,
+    cache_dir: Path | None,
+    force_cpu: bool,
+    detected_device: str,
+    manifest: RecipeManifest,
+    recipe: Recipe,
+    validator: CapabilityValidator | None,
+) -> CapabilityReport:
+    """Prove this node can serve the selection, or refuse to advertise it.
+
+    Runs on the loaded backend before the server starts listening, so a node that
+    cannot execute its shard never reaches a routable endpoint, never registers,
+    and never accepts paid work. `CapabilityAdmissionError` propagates to the CLI,
+    which exits non-zero.
+    """
+    backend = getattr(node, "backend", None)
+    context = CapabilityContext(
+        backend=backend,
+        selection=DoctorSelection(
+            model_id=model_id,
+            shard_start=shard_start,
+            shard_end=shard_end,
+            quantization=quantization,
+            cache_dir=cache_dir,
+            force_cpu=force_cpu,
+        ),
+        recipe=recipe,
+        manifest=manifest,
+        device=_capability_device(backend, detected_device),
+    )
+    print(
+        f"Validating capability — {model_id} layers {shard_start}–{shard_end}, "
+        f"recipe {recipe.id}...",
+        flush=True,
+    )
+    report = (validator or probe_capability)(context)
+    setattr(node, "capability_report", report)  # local evidence, passed or failed
+    admit(AdmissionRequirement.for_context(context), report)
+    print(f"  Capability proven on {context.device} ({report.duration_ms} ms)", flush=True)
+    return report
+
+
 def run_startup(
     tracker_url: str,
     port: int = 0,
@@ -668,6 +740,8 @@ def run_startup(
     torch_interop_threads: int | None = None,
     node_name: str | None = None,
     force_cpu: bool = False,
+    recipe_id: str | None = None,
+    capability_validator: CapabilityValidator | None = None,
 ) -> StubNodeServer | TorchNodeServer:
     """Execute the full startup sequence and return a running node server.
 
@@ -676,13 +750,18 @@ def run_startup(
       2. Load or generate Solana wallet keypair
       3. Query tracker for optimal shard assignment
       4. Download (or stub) the assigned shard from peers, then HuggingFace
-      5. Start local HTTP server
-      6. Register with tracker
+      5. Prove the loaded shard runs — a failure here exits before step 6
+      6. Start local HTTP server and register with tracker
+
+    `capability_validator` is how step 5 is proven. It defaults to a real forward
+    through the loaded shard; only tests replace it, and only with the explicit
+    seams in `meshnet_node.testing` — there is no bypass a deployment can reach.
 
     Prints a compact status summary on completion.
     """
 
     tracker_url = tracker_url.rstrip("/")
+    manifest, recipe = _resolve_recipe(recipe_id)
     relay_url = _discover_relay_url(tracker_url)
     display_fields = _registration_display_fields(node_name)
     if max_loaded_shards < 1:
@@ -874,6 +953,20 @@ def run_startup(
             debug=debug,
             max_loaded_shards=max_loaded_shards,
             force_cpu=force_cpu,
+            recipe_params=recipe.params,
+        )
+        capability_report = _admit_capability(
+            node,
+            model_id=model_id,
+            shard_start=shard_start,
+            shard_end=shard_end,
+            quantization=quantization,
+            cache_dir=cache_dir,
+            force_cpu=force_cpu,
+            detected_device=device,
+            manifest=manifest,
+            recipe=recipe,
+            validator=capability_validator,
         )
         _node_start_time = time.monotonic()
         actual_port = node.start()
@@ -910,6 +1003,11 @@ def run_startup(
             "tracker_mode": (shard_start == 0),
             "managed_assignment": not user_pinned_shard,
             "model_metadata": model_metadata_for(model_id, total_layers, cache_dir=cache_dir),
+            "capability_report": capability_report.to_dict(),
+            # Declared independently of the proof: the tracker checks that the
+            # recipe this node says it serves with is the one the proof ran.
+            "recipe_id": recipe.id,
+            "recipe_version": recipe.version,
             "downloaded_models": (
                 _downloaded_model_inventory(
                     model_id.split("/")[-1],
@@ -1030,6 +1128,20 @@ def run_startup(
             debug=debug,
             max_loaded_shards=max_loaded_shards,
             force_cpu=force_cpu,
+            recipe_params=recipe.params,
+        )
+        capability_report = _admit_capability(
+            node,
+            model_id=assigned_hf_repo,
+            shard_start=assigned_shard_start,
+            shard_end=assigned_shard_end,
+            quantization=quantization,
+            cache_dir=cache_dir,
+            force_cpu=force_cpu,
+            detected_device=device,
+            manifest=manifest,
+            recipe=recipe,
+            validator=capability_validator,
         )
         _node_start_time = time.monotonic()
         actual_port = node.start()
@@ -1062,6 +1174,11 @@ def run_startup(
             "tracker_mode": (assigned_shard_start == 0),
             "managed_assignment": True,
             "model_metadata": model_metadata_for(assigned_hf_repo, assigned_num_layers, cache_dir=cache_dir),
+            "capability_report": capability_report.to_dict(),
+            # Declared independently of the proof: the tracker checks that the
+            # recipe this node says it serves with is the one the proof ran.
+            "recipe_id": recipe.id,
+            "recipe_version": recipe.version,
             "downloaded_models": (
                 _downloaded_model_inventory(
                     assigned_hf_repo.split("/")[-1],
@@ -1212,6 +1329,20 @@ def run_startup(
             debug=debug,
             max_loaded_shards=max_loaded_shards,
             force_cpu=force_cpu,
+            recipe_params=recipe.params,
+        )
+        capability_report = _admit_capability(
+            node,
+            model_id=hf_repo,
+            shard_start=shard_start,
+            shard_end=shard_end,
+            quantization=quantization,
+            cache_dir=shard_path,
+            force_cpu=force_cpu,
+            detected_device=device,
+            manifest=manifest,
+            recipe=recipe,
+            validator=capability_validator,
         )
         actual_port = node.start()
         total_layers = getattr(getattr(node, "backend", None), "total_layers", None) or assigned_total_layers
@@ -1247,6 +1378,11 @@ def run_startup(
             "tracker_mode": (shard_start == 0),
             "managed_assignment": not user_pinned_shard,
             "model_metadata": model_metadata_for(hf_repo, total_layers, cache_dir=shard_path),
+            "capability_report": capability_report.to_dict(),
+            # Declared independently of the proof: the tracker checks that the
+            # recipe this node says it serves with is the one the proof ran.
+            "recipe_id": recipe.id,
+            "recipe_version": recipe.version,
             **registration_capabilities,
             **relay_fields,
             **display_fields,
@@ -1282,6 +1418,19 @@ def run_startup(
         model=assigned_model,
         shard_path=shard_path,
     )
+    capability_report = _admit_capability(
+        node,
+        model_id=assigned_model,
+        shard_start=shard_start,
+        shard_end=shard_end,
+        quantization=quantization,
+        cache_dir=shard_path,
+        force_cpu=force_cpu,
+        detected_device=device,
+        manifest=manifest,
+        recipe=recipe,
+        validator=capability_validator,
+    )
     actual_port = node.start()
     public_host = advertise_host or (socket.getfqdn() if host == "0.0.0.0" else host)
     endpoint = f"http://{public_host}:{actual_port}"
@@ -1304,6 +1453,11 @@ def run_startup(
         "shard_start": shard_start,
         "shard_end": shard_end,
         "shard_checksum": shard_checksum,
+        "capability_report": capability_report.to_dict(),
+        # Declared independently of the proof: the tracker checks that the
+        # recipe this node says it serves with is the one the proof ran.
+        "recipe_id": recipe.id,
+        "recipe_version": recipe.version,
         "downloaded_models": downloaded_models,
         "hardware_profile": hw,
         "wallet_address": address,

packages/node/meshnet_node/testing.py

@@ -0,0 +1,70 @@
+"""Test-only seams. Nothing in the production code path may import this module.
+
+Startup admits a node only on a capability report produced by a *real* forward
+through the loaded shard (see :mod:`meshnet_node.admission`). Tests run against
+fake or stub backends that cannot perform one, so they pass an explicit validator
+from here instead — the honest statement being "this test asserts capability it
+never proved", which is a thing a test may do and a node may not.
+
+`capability_stub` builds the deliberately-wrong reports the fail-closed tests
+need: a failed one, one for another model or shard, one that has aged out.
+"""
+
+from __future__ import annotations
+
+import time
+from typing import Any
+
+from .admission import CapabilityContext, CapabilityValidator
+from .capability import STATUS_PASSED, CapabilityReport, build_capability_report
+
+
+def capability_report_for(
+    context: CapabilityContext,
+    *,
+    status: str = STATUS_PASSED,
+    model_id: str | None = None,
+    shard_start: int | None = None,
+    shard_end: int | None = None,
+    recipe_id: str | None = None,
+    recipe_version: str | None = None,
+    backend_id: str | None = None,
+    device: str | None = None,
+    validated_at: float | None = None,
+    age_seconds: float = 0.0,
+    diagnostics: Any = None,
+    duration_ms: int = 0,
+) -> CapabilityReport:
+    """A report describing `context`, with any field bent away from the truth."""
+    now = time.time() if validated_at is None else validated_at
+    return build_capability_report(
+        model_id=model_id or context.selection.model_id,
+        shard_start=(
+            context.selection.shard_start if shard_start is None else shard_start
+        ),
+        shard_end=context.selection.shard_end if shard_end is None else shard_end,
+        recipe_id=recipe_id or context.recipe.id,
+        recipe_version=recipe_version or context.recipe.version,
+        catalogue_version=context.manifest.catalogue_version,
+        backend_id=backend_id or context.recipe.backend_id,
+        device=device or context.device,
+        quantization=context.selection.quantization,
+        status=status,
+        duration_ms=duration_ms,
+        diagnostics=diagnostics,
+        validated_at=now - age_seconds,
+    )
+
+
+def assume_capability(context: CapabilityContext) -> CapabilityReport:
+    """Assert the selection works, without proving it. Tests only."""
+    return capability_report_for(context)
+
+
+def capability_stub(**overrides: Any) -> CapabilityValidator:
+    """A validator producing a report that deviates from `context` as named."""
+
+    def validator(context: CapabilityContext) -> CapabilityReport:
+        return capability_report_for(context, **overrides)
+
+    return validator

packages/node/meshnet_node/torch_server.py

@@ -3,17 +3,17 @@
 from __future__ import annotations
 
 import base64
+import http.client
 import http.server
 import json
 import sys
 import threading
 import time
-import urllib.error
 import urllib.parse
 import urllib.request
 import uuid
 from pathlib import Path
-from typing import Any
+from typing import Any, Mapping
 
 from .model_backend import (
     InsufficientVRAMError,
@@ -22,16 +22,32 @@ from .model_backend import (
     Quantization,
     TailTokenResult,
     TorchModelShard,
+    _tensor_from_bfloat16_bytes,
     validate_quantization,
 )
+from .seam_telemetry import GenerationTelemetry
+from .activation_compression import (
+    CompressionPolicies,
+    CompressionPolicy,
+    compress_activation,
+    decompress_activation,
+)
 
 
 class _PipelineCacheMiss(Exception):
     """A downstream hop reported 409 cache_miss — head must re-prefill."""
+
+
+class _RelayRequestUncertainError(ConnectionError):
+    """A relay request may have reached the peer but produced no response."""
+
+
+class _DirectRequestUncertainError(ConnectionError):
+    """A direct request may have reached the downstream node but did not finish."""
+
+
 from .server import (
     _WIRE_VERSION,
-    _compress_body,
-    _decompress_body,
     _parse_shape,
     _validate_activation_body,
 )
@@ -107,6 +123,7 @@ class _RelayHopClient:
         self.relay_addr = relay_addr
         self.timeout = timeout
         self._ws = None
+        self._lock = threading.RLock()
 
     def request(
         self,
@@ -118,42 +135,116 @@ class _RelayHopClient:
 
         from .relay_bridge import decode_binary_frame, encode_binary_frame, ws_max_size
 
-        if self._ws is None:
-            self._ws = wsc.connect(
-                self.relay_addr,
-                open_timeout=self.timeout,
-                max_size=ws_max_size(),
-                compression=None,
-            )
-        request_id = f"{time.time_ns():x}"
-        frame = encode_binary_frame({
-            "request_id": request_id,
-            "method": "POST",
-            "path": path,
-            "headers": headers,
-        }, body)
-        self._ws.send(frame)
-        raw = self._ws.recv(timeout=self.timeout)
-        if isinstance(raw, (bytes, bytearray)):
-            resp_header, resp_body = decode_binary_frame(bytes(raw))
-            status = int(resp_header.get("status", 503))
-            resp_headers = {k.lower(): v for k, v in (resp_header.get("headers") or {}).items()}
-            return status, resp_headers, resp_body
-        resp = json.loads(raw)
-        status = int(resp.get("status", 503))
-        resp_headers = {k.lower(): v for k, v in (resp.get("headers") or {}).items()}
-        body_b64 = resp.get("body_base64")
-        resp_body = base64.b64decode(body_b64) if body_b64 else (resp.get("body") or "").encode()
-        return status, resp_headers, resp_body
+        with self._lock:
+            if self._ws is None:
+                self._ws = wsc.connect(
+                    self.relay_addr,
+                    open_timeout=self.timeout,
+                    max_size=ws_max_size(),
+                    compression=None,
+                )
+            request_id = headers.get("X-Meshnet-Activation-Id") or uuid.uuid4().hex
+            frame = encode_binary_frame({
+                "request_id": request_id,
+                "method": "POST",
+                "path": path,
+                "headers": headers,
+            }, body)
+            try:
+                # A send failure is uncertain too: bytes may already have been
+                # accepted by the kernel or peer before the exception surfaced.
+                self._ws.send(frame)
+                raw = self._ws.recv(timeout=self.timeout)
+                if isinstance(raw, (bytes, bytearray)):
+                    resp_header, resp_body = decode_binary_frame(bytes(raw))
+                    response_id = str(resp_header.get("request_id") or "")
+                    status = int(resp_header.get("status", 503))
+                    resp_headers = {
+                        k.lower(): v for k, v in (resp_header.get("headers") or {}).items()
+                    }
+                else:
+                    resp = json.loads(raw)
+                    response_id = str(resp.get("request_id") or "")
+                    status = int(resp.get("status", 503))
+                    resp_headers = {
+                        k.lower(): v for k, v in (resp.get("headers") or {}).items()
+                    }
+                    body_b64 = resp.get("body_base64")
+                    resp_body = (
+                        base64.b64decode(body_b64)
+                        if body_b64 else (resp.get("body") or "").encode()
+                    )
+                if response_id and response_id != request_id:
+                    raise ValueError("relay response request_id did not match request")
+                return status, resp_headers, resp_body
+            except Exception as exc:
+                self.close()
+                raise _RelayRequestUncertainError(
+                    "relay connection failed after forwarding request; refusing replay"
+                ) from exc
 
     def close(self) -> None:
-        if self._ws is not None:
+        with self._lock:
+            if self._ws is not None:
+                try:
+                    self._ws.close()
+                except Exception:
+                    pass
+                finally:
+                    self._ws = None
+
+
+class _DirectHopClient:
+    """One serialized HTTP/1.1 connection to one downstream hop.
+
+    Generation handlers own these clients, so a cached Route Session reuses a
+    TCP connection without sharing it between concurrent Route Sessions.
+    """
+
+    def __init__(self, endpoint: str, timeout: float = 120.0) -> None:
+        parsed = urllib.parse.urlsplit(endpoint.rstrip("/"))
+        if parsed.scheme not in {"http", "https"} or not parsed.hostname:
+            raise ValueError(f"invalid downstream endpoint: {endpoint!r}")
+        self.timeout = timeout
+        self._scheme = parsed.scheme
+        self._host = parsed.hostname
+        self._port = parsed.port
+        self._base_path = parsed.path.rstrip("/")
+        self._connection: http.client.HTTPConnection | None = None
+        self._lock = threading.RLock()
+
+    def _connect(self) -> http.client.HTTPConnection:
+        connection_type = (
+            http.client.HTTPSConnection if self._scheme == "https" else http.client.HTTPConnection
+        )
+        return connection_type(self._host, self._port, timeout=self.timeout)
+
+    def request(
+        self, path: str, body: bytes, headers: Mapping[str, str],
+    ) -> tuple[int, dict[str, str], bytes]:
+        request_path = f"{self._base_path}{path if path.startswith('/') else '/' + path}"
+        with self._lock:
+            if self._connection is None:
+                self._connection = self._connect()
             try:
-                self._ws.close()
-            except Exception:
-                pass
-            finally:
-                self._ws = None
+                self._connection.request("POST", request_path, body=body, headers=dict(headers))
+                response = self._connection.getresponse()
+                response_body = response.read()
+                response_headers = {key.lower(): value for key, value in response.headers.items()}
+                return response.status, response_headers, response_body
+            except Exception as exc:
+                self.close()
+                raise _DirectRequestUncertainError(
+                    "direct connection failed after forwarding request; refusing replay"
+                ) from exc
+
+    def close(self) -> None:
+        with self._lock:
+            if self._connection is not None:
+                try:
+                    self._connection.close()
+                finally:
+                    self._connection = None
 
 
 def _relay_hop(
@@ -178,18 +269,15 @@ def _relay_hop(
         client.close()
 
 
-# Below this, zstd overhead outweighs the win (per-token decode bodies are ~KBs).
-_COMPRESS_MIN_BYTES = 64 * 1024
+_COMPRESSION_POLICIES = CompressionPolicies()
 
 
-def _maybe_compress_activation(body: bytes) -> tuple[bytes, str | None]:
-    """zstd-compress large activation bodies; returns (wire_body, encoding)."""
-    if len(body) < _COMPRESS_MIN_BYTES:
-        return body, None
-    try:
-        return _compress_body(body, "zstd"), "zstd"
-    except Exception:
-        return body, None
+def _maybe_compress_activation(
+    body: bytes, policy: CompressionPolicy | None = None,
+) -> tuple[bytes, str | None]:
+    """Compatibility wrapper for callers that only need wire body and encoding."""
+    result = compress_activation(body, policy or _COMPRESSION_POLICIES.for_condition("lan", "prefill"))
+    return result.body, result.encoding
 
 
 def _is_cache_miss_body(body: bytes) -> bool:
@@ -285,6 +373,7 @@ class _TorchHTTPServer(http.server.HTTPServer):
                     "elapsed_seconds": round(elapsed, 1),
                     "tokens_per_sec": round(tokens / elapsed, 2) if tokens > 0 else 0.0,
                     "routing_complete": bool(rec.get("routing_complete")),
+                    "telemetry": rec["telemetry"].snapshot(now=now) if rec.get("telemetry") else None,
                 })
             return out
 
@@ -306,6 +395,10 @@ class _TorchHTTPServer(http.server.HTTPServer):
 
 
 class _TorchHandler(http.server.BaseHTTPRequestHandler):
+    # HTTP/1.1 is required for Route Session-owned downstream connections.
+    # Finite responses below provide Content-Length; streams are chunked.
+    protocol_version = "HTTP/1.1"
+
     def log_message(self, fmt, *args):  # noqa: suppress request logs in tests
         pass
 
@@ -317,6 +410,9 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         )
 
     def _request_log_suffix(self) -> str:
+        activation_id = self.headers.get("X-Meshnet-Activation-Id")
+        if activation_id:
+            return f" activation_id={activation_id}"
         req_id = self.headers.get("X-Meshnet-Request-Id") or self.headers.get("X-Request-Id")
         return f" request_id={req_id}" if req_id else ""
 
@@ -334,6 +430,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 "started": time.monotonic(),
                 "tokens": 0,
                 "routing_complete": False,
+                "telemetry": None,
             }
 
     def _track_request_progress(
@@ -366,6 +463,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             self._handle_chat_completions()
         else:
             self.send_response(404)
+            self.send_header("Content-Length", "0")
             self.end_headers()
 
     def _handle_infer(self) -> None:
@@ -427,7 +525,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             encoding = self.headers.get("X-Meshnet-Encoding")
             length = int(self.headers.get("Content-Length", 0))
             body = self.rfile.read(length)
-            raw_body = _decompress_body(body, encoding)
+            raw_body = decompress_activation(body, encoding).body
             _validate_activation_body(raw_body, shape, dtype)
             if dtype != "bfloat16":
                 raise ValueError("real model backend requires bfloat16 activation input")
@@ -438,6 +536,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         except (KeyError, ValueError, TypeError):
             self.send_response(400)
             self.send_header("X-Meshnet-Wire", _WIRE_VERSION)
+            self.send_header("Content-Length", "0")
             self.end_headers()
             return
 
@@ -511,7 +610,12 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             self._send_json(200, {"text": result})
             return
 
-        response_body = _compress_body(result.body, encoding)
+        route_condition = self.headers.get("X-Meshnet-Compression-Route", "lan")
+        phase_condition = cache_mode if cache_mode in {"prefill", "decode"} else "prefill"
+        response_compression = compress_activation(
+            result.body, _COMPRESSION_POLICIES.for_condition(route_condition, phase_condition),
+        )
+        response_body = response_compression.body
         self.send_response(200)
         self.send_header("Content-Type", "application/octet-stream")
         self.send_header("Content-Length", str(len(response_body)))
@@ -521,8 +625,8 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         self.send_header("X-Meshnet-Session", session)
         self.send_header("X-Meshnet-Chunk-Index", chunk_index)
         self.send_header("X-Meshnet-Chunk-Total", chunk_total)
-        if encoding:
-            self.send_header("X-Meshnet-Encoding", encoding)
+        if response_compression.encoding:
+            self.send_header("X-Meshnet-Encoding", response_compression.encoding)
         if result.attention_mask_header:
             self.send_header("X-Meshnet-Attn-Mask", result.attention_mask_header)
         if result.position_ids_header:
@@ -700,6 +804,11 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         current_text = prompt_text
 
         session_id = str(uuid.uuid4())
+        telemetry = GenerationTelemetry(session_id)
+        with server._stats_lock:
+            current = server._active_requests.get(request_id)
+            if current is not None:
+                current["telemetry"] = telemetry
         use_kv = bool(getattr(backend, "supports_kv_cache", False))
         # EOS detection by id must work on the stateless path too: the tail
         # returns token_id regardless of caching, and EOS usually decodes to
@@ -722,6 +831,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         last_token_id: int | None = None
         failure_reason: str | None = None
         relay_clients: dict[str, _RelayHopClient] = {}
+        direct_clients: dict[str, _DirectHopClient] = {}
 
         def _prefill_step() -> tuple[str, int | None]:
             """Full-sequence prefill: initial step and cache-miss recovery."""
@@ -734,6 +844,8 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 payload, remaining_route, backend=backend,
                 session=session_id, cache_mode="prefill" if use_kv else None,
                 relay_clients=relay_clients,
+                direct_clients=direct_clients if use_kv else None,
+                telemetry=telemetry,
             )
 
         for step in range(max_tokens):
@@ -745,6 +857,8 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                             payload, remaining_route, backend=backend,
                             session=session_id, cache_mode="decode",
                             relay_clients=relay_clients,
+                            direct_clients=direct_clients,
+                            telemetry=telemetry,
                         )
                     except (KVCacheMiss, _PipelineCacheMiss) as miss:
                         # Evicted/restarted node or head lost its own session:
@@ -758,11 +872,11 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 else:
                     token_str, token_id = _prefill_step()
             except _PipelineCacheMiss as exc:
-                print(f"  [node] unexpected cache miss on prefill: {exc}", flush=True)
+                print(f"  [node] unexpected cache miss on prefill session={session_id[:8]}: {exc}", flush=True)
                 failure_reason = f"cache miss on prefill: {exc}"
                 break
             except Exception as exc:
-                print(f"  [node] distributed encode error: {exc}", flush=True)
+                print(f"  [node] distributed encode error session={session_id[:8]}: {exc}", flush=True)
                 failure_reason = f"distributed encode error: {exc}"
                 break
             # Stop on error responses or EOS.
@@ -789,14 +903,32 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 tokens=len(generated),
                 routing_complete=True,
             )
+            telemetry.note_tokens(len(generated))
             now = time.monotonic()
+            if telemetry.report_due:
+                summary = telemetry.snapshot(now=now)
+                seams = summary["seams"]
+                if seams:
+                    latest = seams[-1]
+                    print(
+                        f"  [node] seam telemetry session={session_id[:8]} "
+                        f"phase={latest['phase']} hop={latest['hop']} "
+                        f"activations={latest['activations']} "
+                        f"avg_ms={latest['avg_latency_ms']:.2f} "
+                        f"wire_bytes={latest['wire_bytes']} "
+                        f"response_bytes={latest['response_bytes']} "
+                        f"tps={summary['tokens_per_sec']:.2f} "
+                        f"activation_id={latest['last_activation_id']}",
+                        flush=True,
+                    )
+                telemetry.mark_reported(now=now)
             if step == 0 or now - last_gen_log >= _GENERATION_LOG_INTERVAL:
                 elapsed = now - gen_started
                 token_count = len(generated)
                 tps = token_count / max(elapsed, 1e-6)
                 _write_progress_line(
                     progress_line,
-                    f"  [node] generating step={step + 1}/{max_tokens} "
+                    f"  [node] generating step={step + 1}/{max_tokens} session={session_id[:8]} "
                     f"tokens={token_count} elapsed_s={elapsed:.1f} tps={tps:.2f}",
                 )
                 last_gen_log = now
@@ -808,6 +940,8 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 pass
         for relay_client in relay_clients.values():
             relay_client.close()
+        for direct_client in direct_clients.values():
+            direct_client.close()
 
         if generated:
             elapsed = time.monotonic() - gen_started
@@ -815,10 +949,11 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             tps = token_count / max(elapsed, 1e-6)
             _write_progress_line(
                 progress_line,
-                f"  [node] generation complete tokens={token_count} "
+                f"  [node] generation complete session={session_id[:8]} tokens={token_count} "
                 f"elapsed_s={elapsed:.1f} tps={tps:.2f}",
                 final=True,
             )
+        telemetry.close()
 
         result_text = "".join(generated)
         # A failure before the first token is an upstream error, not an empty
@@ -921,6 +1056,8 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         session: str | None = None,
         cache_mode: str | None = None,
         relay_clients: dict[str, _RelayHopClient] | None = None,
+        direct_clients: dict[str, _DirectHopClient] | None = None,
+        telemetry: GenerationTelemetry | None = None,
     ) -> tuple[str, int | None]:
         """Forward an activation through the downstream route.
 
@@ -935,10 +1072,9 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             # Full single-node (head+tail) is handled before entering this method.
             if active_backend.is_tail:
                 try:
-                    tensor = active_backend.torch.frombuffer(
-                        bytearray(payload.body),  # type: ignore[union-attr]
-                        dtype=active_backend.torch.bfloat16,
-                    ).reshape(payload.shape).to(active_backend.device)  # type: ignore[union-attr]
+                    tensor = _tensor_from_bfloat16_bytes(
+                        payload.body, payload.shape, active_backend.torch,  # type: ignore[union-attr]
+                    ).to(active_backend.device)
                     if hasattr(active_backend, "decode_tail_token"):
                         tail = active_backend.decode_tail_token(tensor)
                         return tail.text, tail.token_id
@@ -968,7 +1104,12 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                     + (f" relay={relay_addr}" if relay_addr else ""),
                     flush=True,
                 )
-            wire_body, wire_encoding = _maybe_compress_activation(current_body)
+            phase = cache_mode if cache_mode in {"prefill", "decode"} else "prefill"
+            route_condition = "relay" if relay_addr else "lan"
+            compression = compress_activation(
+                current_body, _COMPRESSION_POLICIES.for_condition(route_condition, phase),
+            )
+            wire_body, wire_encoding = compression.body, compression.encoding
             headers: dict[str, str] = {
                 "Content-Type": "application/octet-stream",
                 "X-Meshnet-Wire": _WIRE_VERSION,
@@ -979,9 +1120,17 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 "X-Meshnet-Chunk-Total": "1",
                 "X-Meshnet-Hop-Index": str(hop_index),
                 "X-Meshnet-Start-Layer": str(start_layer),
+                "X-Meshnet-Activation-Id": uuid.uuid4().hex,
+                "X-Meshnet-Compression-Route": route_condition,
             }
             if wire_encoding:
                 headers["X-Meshnet-Encoding"] = wire_encoding
+            if telemetry is not None:
+                telemetry.record_compression(
+                    phase=phase, hop=hop_index, node=node_url,
+                    input_bytes=compression.input_bytes, output_bytes=compression.output_bytes,
+                    elapsed_seconds=compression.elapsed_seconds,
+                )
             if cache_mode:
                 headers["X-Meshnet-Cache"] = cache_mode
                 past_len = getattr(payload, "past_len", None)
@@ -992,6 +1141,10 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             if current_pos:
                 headers["X-Meshnet-Position-Ids"] = current_pos
             if relay_addr:
+                connection_reused = bool(
+                    relay_clients is not None and relay_addr in relay_clients
+                )
+                seam_started = time.monotonic()
                 try:
                     if relay_clients is None:
                         status, resp_headers, resp_body = _relay_hop(
@@ -1004,44 +1157,100 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                         status, resp_headers, resp_body = relay_client.request(
                             "/forward", wire_body, headers,
                         )
+                    if telemetry is not None:
+                        telemetry.record_seam(
+                            activation_id=headers["X-Meshnet-Activation-Id"],
+                            phase=cache_mode or "prefill",
+                            hop=hop_index,
+                            node=node_url,
+                            latency_seconds=time.monotonic() - seam_started,
+                            wire_bytes=len(wire_body),
+                            response_bytes=len(resp_body),
+                            connection_reused=connection_reused,
+                        )
                     if status == 409 and _is_cache_miss_body(resp_body):
                         raise _PipelineCacheMiss(node_url)
                     if status >= 400:
                         detail = _response_error_snippet(resp_body)
                         print(
-                            f"  [node] relay hop {hop_index} returned {status} from {relay_addr}: {detail}",
+                            f"  [node] relay hop {hop_index} session={session[:8]} returned "
+                            f"{status} from {relay_addr}: {detail}",
                             flush=True,
                         )
                         return f"pipeline error at {node_url} via relay: status {status}: {detail}", None
                 except _PipelineCacheMiss:
                     raise
+                except _RelayRequestUncertainError as exc:
+                    # The activation may already have mutated the downstream
+                    # KV cache.  Do not replay it on a direct connection.
+                    print(
+                        f"  [node] relay hop {hop_index} session={session[:8]} outcome is uncertain at "
+                        f"{relay_addr}: {exc}",
+                        flush=True,
+                    )
+                    return f"pipeline relay outcome uncertain at {node_url}: {exc}", None
                 except Exception as exc:
                     print(
-                        f"  [node] relay hop {hop_index} failed at {relay_addr}: {exc}; "
+                        f"  [node] relay hop {hop_index} session={session[:8]} failed at {relay_addr}: {exc}; "
                         f"falling back to direct {node_url}",
                         flush=True,
                     )
                     relay_addr = None  # fall through to direct
             if not relay_addr:
-                req = urllib.request.Request(
-                    f"{node_url}/forward",
-                    data=wire_body,
-                    headers=headers,
-                    method="POST",
+                connection_reused = bool(
+                    direct_clients is not None and node_url in direct_clients
                 )
+                seam_started = time.monotonic()
                 try:
-                    with urllib.request.urlopen(req, timeout=120.0) as r:
-                        resp_body = r.read()
-                        resp_headers = {k.lower(): v for k, v in r.headers.items()}
-                except urllib.error.HTTPError as exc:
-                    body = exc.read()
-                    if exc.code == 409 and _is_cache_miss_body(body):
-                        raise _PipelineCacheMiss(node_url) from exc
-                    detail = _response_error_snippet(body)
-                    print(f"  [node] pipeline hop {hop_index} failed at {node_url}: {exc}: {detail}", flush=True)
-                    return f"pipeline error at {node_url}: {exc}: {detail}", None
+                    if direct_clients is None:
+                        direct_client = _DirectHopClient(node_url, timeout=120.0)
+                        try:
+                            status, resp_headers, resp_body = direct_client.request(
+                                "/forward", wire_body, headers,
+                            )
+                        finally:
+                            direct_client.close()
+                    else:
+                        direct_client = direct_clients.setdefault(
+                            node_url, _DirectHopClient(node_url, timeout=120.0),
+                        )
+                        status, resp_headers, resp_body = direct_client.request(
+                            "/forward", wire_body, headers,
+                        )
+                    if telemetry is not None:
+                        telemetry.record_seam(
+                            activation_id=headers["X-Meshnet-Activation-Id"],
+                            phase=cache_mode or "prefill",
+                            hop=hop_index,
+                            node=node_url,
+                            latency_seconds=time.monotonic() - seam_started,
+                            wire_bytes=len(wire_body),
+                            response_bytes=len(resp_body),
+                            connection_reused=connection_reused,
+                        )
+                    if status == 409 and _is_cache_miss_body(resp_body):
+                        raise _PipelineCacheMiss(node_url)
+                    if status >= 400:
+                        detail = _response_error_snippet(resp_body)
+                        print(
+                            f"  [node] pipeline hop {hop_index} session={session[:8]} failed at {node_url}: "
+                            f"status {status}: {detail}", flush=True,
+                        )
+                        return f"pipeline error at {node_url}: status {status}: {detail}", None
+                except _PipelineCacheMiss:
+                    raise
+                except _DirectRequestUncertainError as exc:
+                    print(
+                        f"  [node] pipeline hop {hop_index} session={session[:8]} outcome is uncertain "
+                        f"at {node_url}: {exc}",
+                        flush=True,
+                    )
+                    return f"pipeline direct outcome uncertain at {node_url}: {exc}", None
                 except Exception as exc:
-                    print(f"  [node] pipeline hop {hop_index} failed at {node_url}: {exc}", flush=True)
+                    print(
+                        f"  [node] pipeline hop {hop_index} session={session[:8]} "
+                        f"failed at {node_url}: {exc}", flush=True,
+                    )
                     return f"pipeline error at {node_url}: {exc}", None
             content_type = resp_headers.get("content-type", "")
             if "application/json" in content_type:
@@ -1058,11 +1267,21 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             shape_header = resp_headers.get("x-meshnet-shape", ",".join(str(d) for d in current_shape))
             current_shape = _parse_shape(shape_header)
             try:
-                current_body = _decompress_body(
+                decompression = decompress_activation(
                     resp_body, resp_headers.get("x-meshnet-encoding")
                 )
+                current_body = decompression.body
+                if telemetry is not None:
+                    telemetry.record_compression(
+                        phase=phase, hop=hop_index, node=node_url,
+                        input_bytes=decompression.input_bytes, output_bytes=decompression.output_bytes,
+                        elapsed_seconds=decompression.elapsed_seconds, decompression=True,
+                    )
             except ValueError as exc:
-                print(f"  [node] pipeline hop {hop_index} bad response encoding: {exc}", flush=True)
+                print(
+                    f"  [node] pipeline hop {hop_index} session={session[:8]} "
+                    f"bad response encoding: {exc}", flush=True,
+                )
                 return f"pipeline error at {node_url}: {exc}", None
             current_attn = resp_headers.get("x-meshnet-attn-mask")
             current_pos = resp_headers.get("x-meshnet-position-ids")
@@ -1084,14 +1303,17 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         self.send_response(200)
         self.send_header("Content-Type", "text/event-stream; charset=utf-8")
         self.send_header("Cache-Control", "no-cache")
+        self.send_header("Transfer-Encoding", "chunked")
         self.end_headers()
 
-        def _emit(data: str) -> None:
+        def _emit(data: str) -> bool:
             try:
-                self.wfile.write(f"data: {data}\n\n".encode())
+                body = f"data: {data}\n\n".encode()
+                self.wfile.write(f"{len(body):X}\r\n".encode() + body + b"\r\n")
                 self.wfile.flush()
+                return True
             except (BrokenPipeError, ConnectionResetError):
-                pass
+                return False
 
         _emit(json.dumps({
             "id": chunk_id, "object": "chat.completion.chunk", "created": created,
@@ -1105,7 +1327,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                 # instead of showing an empty completion.
                 _emit(json.dumps({"error": {"message": error, "type": "upstream_error"}}))
                 try:
-                    self.wfile.write(b"data: [DONE]\n\n")
+                    self.wfile.write(b"E\r\ndata: [DONE]\n\n\r\n0\r\n\r\n")
                     self.wfile.flush()
                 except (BrokenPipeError, ConnectionResetError):
                     pass
@@ -1117,7 +1339,7 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
                     "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
                 }))
                 try:
-                    self.wfile.write(b"data: [DONE]\n\n")
+                    self.wfile.write(b"E\r\ndata: [DONE]\n\n\r\n0\r\n\r\n")
                     self.wfile.flush()
                 except (BrokenPipeError, ConnectionResetError):
                     pass
@@ -1159,14 +1381,17 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
         self.send_response(200)
         self.send_header("Content-Type", "text/event-stream; charset=utf-8")
         self.send_header("Cache-Control", "no-cache")
+        self.send_header("Transfer-Encoding", "chunked")
         self.end_headers()
 
-        def _emit(data: str) -> None:
+        def _emit(data: str) -> bool:
             try:
-                self.wfile.write(f"data: {data}\n\n".encode())
+                body = f"data: {data}\n\n".encode()
+                self.wfile.write(f"{len(body):X}\r\n".encode() + body + b"\r\n")
                 self.wfile.flush()
-            except BrokenPipeError:
-                pass
+                return True
+            except (BrokenPipeError, ConnectionResetError):
+                return False
 
         _emit(json.dumps({
             "id": chunk_id, "object": "chat.completion.chunk", "created": created,
@@ -1184,9 +1409,9 @@ class _TorchHandler(http.server.BaseHTTPRequestHandler):
             "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
         }))
         try:
-            self.wfile.write(b"data: [DONE]\n\n")
+            self.wfile.write(b"E\r\ndata: [DONE]\n\n\r\n0\r\n\r\n")
             self.wfile.flush()
-        except BrokenPipeError:
+        except (BrokenPipeError, ConnectionResetError):
             pass
 
 
@@ -1255,6 +1480,7 @@ class TorchNodeServer:
         debug: bool = False,
         max_loaded_shards: int = 1,
         force_cpu: bool = False,
+        recipe_params: Mapping[str, Any] | None = None,
     ) -> None:
         self._host = host
         self._requested_port = port
@@ -1266,6 +1492,7 @@ class TorchNodeServer:
             quantization,
             cache_dir,
             force_cpu=force_cpu,
+            recipe_params=recipe_params,
         )
         self._backends: dict[str, TorchModelShard] = {self._backend.model_id: self._backend}
         # Auto-detect tracker mode: enabled when shard_start == 0 or explicitly set
@@ -1415,13 +1642,20 @@ def _load_backend(
     quantization: str,
     cache_dir: Path | None = None,
     force_cpu: bool = False,
+    recipe_params: Mapping[str, Any] | None = None,
 ) -> TorchModelShard:
     from .model_backend import load_torch_shard
 
     quant = validate_quantization(quantization)
     try:
         return load_torch_shard(
-            model_id, shard_start, shard_end, quant, cache_dir, force_cpu=force_cpu
+            model_id,
+            shard_start,
+            shard_end,
+            quant,
+            cache_dir,
+            force_cpu=force_cpu,
+            recipe_params=recipe_params,
         )
     except MissingModelDependencyError:
         raise