fix(cost): write ledger records with json.dumps so quoted model ids aren't dropped

[RealDiligent]

Problem

_ledger_append (src/trinity/llm/openai_compatible_pool.py) builds each cost-ledger line by hand:

f.write(
    f'{{"provider":"{provider}","m":"{model}","p":{int(prompt_tokens)},'
    f'"c":{int(completion_tokens)}}}\n'
)

model is route.model_id, read verbatim from configs/models*.yaml. A model id containing a double quote ("), backslash (\), or control character makes the line invalid JSON. The reader, ledger_cost_report (src/trinity/costing.py), wraps json.loads in try/except json.JSONDecodeError: continue, so those lines are silently skipped.

Impact

Silent under-reporting of API spend: every call for such a model disappears from cost_usd, cost_calls, and the per-model breakdown, with no error — and the JSONL file is corrupted for any other consumer. The cost ledger exists to prevent exactly this kind of drift.

Fix

Serialize the record with json.dumps so the writer and reader agree on escaping:

-        with open(path, "a") as f:
-            f.write(
-                f'{{"provider":"{provider}","m":"{model}","p":{int(prompt_tokens)},'
-                f'"c":{int(completion_tokens)}}}\n'
-            )
+        record = json.dumps(
+            {"provider": provider, "m": model, "p": int(prompt_tokens), "c": int(completion_tokens)}
+        )
+        with open(path, "a") as f:
+            f.write(record + "\n")

The line format (keys provider/m/p/c, one JSON object per line) is unchanged for well-behaved ids, so existing ledgers keep parsing.

Tests

New offline tests/test_cost_ledger.py round-trips _ledger_append → ledger_cost_report, including a model id containing a quote and a backslash. Verified the metacharacter test fails on the old f-string (cost_calls == 0, call dropped) and passes with the fix. Full router suite green locally (only the pre-existing test_run_remote / test_setup_remote python3-on-PATH failures remain, unrelated).

Risk / tradeoffs

Minimal. Same on-disk format for normal ids; only adds correct escaping. No API or config changes.

Closes #165

[RealDiligent]

Note on CI: the test-router failure here is pre-existing on main, not from this change. It's tests/test_validate_submission.py::test_wrong_theta_length_fails, which went stale after #150 made the submission size check warning-only — main's own CI is currently red on the same test (tracked in #160, fix open in #161). This PR touches only the cost ledger (openai_compatible_pool.py / costing.py); test-validator and web are green, and the new test_cost_ledger.py passes. Once #161 lands, a re-run here goes green.

[RealDiligent]

Updated: rebased onto current main and CI is now green (test-router, test-validator, web all pass).

The earlier red here was not from the cost-ledger change — it was the pre-existing test_wrong_theta_length_fails, which went dead after the submission size check became warning-only (main's own CI is red on it for every branch; see #160). This PR now also un-stales that test (rename → test_wrong_theta_length_warns, asserting the current warning-only behavior) so CI passes. Happy to drop that commit if #160/#161 lands first — the cost-ledger fix (json.dumps in _ledger_append) is the substantive change and is independent.