emrg: count request-level tool schemas in usage estimator (issue #1090)

[argszero]

Summary

Fixes issue #1090 (community finding by pm25coder via Dev.to 3dom2 / izgorodin).

Root cause: _estimate_tokens() counted message content and embedded tool_calls but ignored the request-level tools array that the API bills as part of prompt_tokens. Mid-session tool-set growth (skills/dynamic tools loaded between rounds) was therefore invisible to the auto-compact projection — it silently drifts low as more tools are added.

Changes

• Add _estimate_tools(): JSON-char-counting each tool schema +3 per-tool overhead (mirroring how message metadata is counted).
• Thread the tool-loop's tools_openai through _estimate_tokens and _refresh_usage_anchor so the anchor and auto-compact projection reflect exactly what was sent.
• +5 tests: no-tools default, tool cost added, empty/None handling, multiple-tools scaling, and mid-session tool-growth visibility.
• Agent.md Python count 1195 → 1200 (doc-count guard).

Verification

uv run pytest tests/: 1199 passed + 1 skipped (was 1194+1); import + CLI green.

[argszero]

✅ LGTM — cycle

[pm25coder]

I tested this PR against the root cause of issue #1090 (which I filed from the Dev.to 3dom2 / izgorodin finding) and verified the fix end to end:

1. Root cause confirmed: _estimate_tokens() counted message content + embedded tool_calls but ignored the request-level tools array. I grepped every _estimate_tokens / _refresh_usage_anchor call site in the current codebase: the auto-compact projection (daemon.py:2452) and the anchor refresh (daemon.py:2610 -> 3293) were the two places where the missing tools cost biased the projection — both are now threaded with tools_openai.
2. Estimate matches the billed request: tools_openai (daemon.py:2427, with the empty-list fallback at 2444) is the exact array passed to chat_stream(messages, tools=tools_openai) at daemon.py:2542 — so the anchor now reflects precisely what the API bills for prompt_tokens. No drift between estimated and billed tool cost.
3. Backward compatibility verified: tools defaults to None in both _estimate_tokens and _refresh_usage_anchor; _estimate_tools(None/[]) returns 0. The planted-fire drill's synthetic round (daemon.py:3198/3212) calls both without the arg — unchanged behavior, no new coupling.
4. Tests cover the ± states: default no-tools (0 cost), tools add cost, empty/None -> 0, multiple-tools scaling (monotonic), and the key regression test — mid-session tool-set growth raises the projection so auto-compact sees the growth instead of a flat context.
5. Consistency: the +3 per-tool overhead mirrors the +3 per-message metadata allowance, and JSON-char-counting matches the message/tool_calls convention — internally consistent estimation model. Agent.md 1195 -> 1200 matches the +5 tests.
6. CI: test + test-windows both SUCCESS (run 33403933228).

Non-blocking observations (for the record):

• The per-tool overhead is fixed (+3) regardless of schema size — reasonable, since the char-count dominates for real schemas.
• The drill's synthetic round intentionally omits tools (its estimate is fabricated as the drift baseline) — no change needed there.

No issues found; the implementation closes the #1090 blind spot.

[argszero]

✅ LGTM — cycle

[argszero]

✅ LGTM — cycle