dual billing; tracker to node model sharing

docs/issues/44-tracker-shard-source-partial-download.md

@@ -1,6 +1,6 @@
 # US-044 — Tracker as model-file source; nodes download only their shard
 
-Status: planned
+Status: in progress
 Priority: High (blocks multi-machine big-model serving; pairs with US-042)
 Stage: Designed (grill remaining decisions before build)
 
@@ -56,10 +56,30 @@ What exists already (build on it, don't duplicate):
 
 ## Acceptance criteria
 
-- Two-machine test: machine A (tracker + node, holds full snapshot) serves
+- [x] Tracker can be started with `--models-dir PATH` / `MESHNET_MODELS_DIR`
+  and advertises a local model-file source in assignment responses when it has
+  a matching HF snapshot.
+- [x] Tracker serves a tar stream containing only the safetensors files selected
+  for the assigned layer range plus config/tokenizer/index metadata.
+- [x] Node downloader keeps exact-shard peers first, then races tracker model
+  sources against a HuggingFace `snapshot_download(..., allow_patterns=...)`
+  subset download, using the first successful source.
+- [x] Real PyTorch model startup can use tracker `full_url` sources to fetch
+  the full local snapshot over LAN before `from_pretrained`, so local-network
+  testing no longer has to pull from HuggingFace first.
+- [ ] Two-machine test: machine A (tracker + node, holds full snapshot) serves
   layers 0–k; machine B joins with no model and receives **only** the files
   for its assigned range from A — nothing fetched from HF
-- Machine B's resident memory scales with its shard size, not model size
-- Checksums verified end-to-end; corrupted transfer falls back cleanly
-- Single-node/full-model flows unchanged
-- `python -m pytest` passes from repo root
+- [ ] Machine B's resident memory scales with its shard size, not model size
+- [ ] Checksums verified end-to-end; corrupted transfer falls back cleanly
+- [x] Single-node/full-model flows unchanged
+- [ ] `python -m pytest` passes from repo root
+
+## Implementation notes
+
+- 2026-07-06: Added the tracker/node download path. For immediate Qwen3.6-35B
+  LAN testing, real PyTorch nodes fetch the full snapshot from the tracker via
+  `full_url` and race HuggingFace as fallback. Remaining hard half is true
+  partial model materialization: the backend can prefer a downloaded local
+  model directory, but Transformers still needs a `meta`-device load path that
+  materializes only assigned layers.

docs/issues/45-dual-rate-billing.md

@@ -0,0 +1,60 @@
+# US-045 — Dual-rate billing: separate input and output token prices
+
+Status: in progress
+Priority: High (billing correctness before friends test; providers all price this way)
+Stage: Designed
+
+## Context
+
+Today the ledger has one `price_per_1k_tokens` per model, and the two proxy
+paths don't even agree on what they count:
+
+- **Non-streaming** bills `usage.total_tokens` (prompt + completion) at the
+  blended rate (`_billable_non_stream_tokens`).
+- **Streaming** bills `min(observed output deltas, reported total)` — output
+  only in practice (`_billable_stream_tokens`).
+- The HF pricing refresher (issue 23) averages a provider's input/output
+  rates 50/50 (`blended_price_per_1k_tokens`), which misprices asymmetric
+  models — e.g. Qwen3.6-35B-A3B on deepinfra is $0.15/1M in, $0.95/1M out.
+
+Decision (user, 2026-07-06): charge **both** input and output tokens, at
+**two separate rates**, same as other providers.
+
+## Design
+
+1. **`BillingLedger`** stores `{model: (input_per_1k, output_per_1k)}`.
+   - `set_prices(model, input_per_1k, output_per_1k)` (new);
+     `set_price(model, p)` keeps working and sets both.
+   - `prices_for(model) -> (input, output)` (new); `price_for(model)` returns
+     the blended average for back-compat (estimators/logs).
+   - `charge_request(...)` gains keyword `input_tokens`/`output_tokens`; when
+     provided, `cost = in_rate·in/1k + out_rate·out/1k` and the event records
+     the split. Without them, legacy behavior (blended × total) — old events
+     and gossip replicas replay unchanged (`cost` stays the applied field).
+2. **Token counting** (`server.py`):
+   - Non-stream: prefer `usage.prompt_tokens`/`completion_tokens`; fall back
+     to content estimates (`_estimate_prompt_tokens`, observed completion),
+     capped by `max_tokens` bounds as today.
+   - Stream (direct + relay): output = observed deltas as today; input =
+     `usage.prompt_tokens` when a usage chunk appears, else the prompt
+     estimate from the request body. `_stream_line_tokens` returns the parsed
+     usage triple instead of just the total.
+3. **Presets**: `input_price_per_1k_tokens` / `output_price_per_1k_tokens`
+   (dual keys win; `price_per_1k_tokens` alone still means "both rates").
+   Qwen3.6-35B-A3B: input 0.00012, output 0.00076 (80% of deepinfra).
+4. **HF refresher**: applies 80% of each side separately via `set_prices`
+   (all alias keys); change log keeps recording the blended pair for history
+   continuity.
+5. **Spend cap** (`--max-charge-per-request`): estimate =
+   `in_rate·prompt_estimate + out_rate·completion_limit`.
+
+## Acceptance criteria
+
+- Streamed and non-streamed requests for the same exchange bill the same
+  split (input charged in both)
+- A model with asymmetric provider rates bills input and output differently;
+  `usage_for` / billing events expose the split
+- Old persisted billing events replay byte-identically (balances unchanged)
+- HF refresh sets both rates from the marketplace row, not the average
+- Spend cap uses the dual rates
+- `python -m pytest` passes from repo root