fix(cost): serialize ledger records with json.dumps so ids can't corrupt JSON (#165)

[minion1227]

Fixes #165

Problem

_ledger_append in src/trinity/llm/openai_compatible_pool.py built each cost-ledger line by hand with an f-string:

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

model is route.model_id, taken verbatim from configs/models*.yaml. If an id contains a double quote, a backslash, or a control character, the emitted line is not valid JSON:

{"provider":"openrouter","m":"weird"model","p":1000,"c":2000}   <- invalid
{"provider":"openrouter","m":"back\slash","p":10,"c":20}        <- invalid escape

costing.ledger_cost_report does try: json.loads(line) except JSONDecodeError: continue, so those lines are silently skipped — their tokens never reach cost_usd, cost_calls, or the per-model breakdown. That is silent under-reporting of API spend, with no error and no warning, and it corrupts the JSONL file for any other consumer.

Reproduced before the fix: appending 3 records where 2 have metacharacter ids gives cost_calls == 1.

Fix

Serialize the record with json.dumps(..., ensure_ascii=False) instead of hand-formatting, so the writer and reader agree on escaping. ensure_ascii=False keeps non-ASCII ids human-readable in the ledger.

After the fix the same 5 records all parse, and cost_calls == 5 with exact token totals and ids preserved (weird"model, back\slash, ctrl\tchar, uniécode/model).

Tests

Adds tests/test_cost_ledger_json.py (14 cases, pure stdlib — no network/GPU/torch):

• quote / backslash / tab / newline / unicode ids each round-trip as valid JSON with the id preserved
• the provider field is escaped too
• no call is dropped from cost_calls / cost_prompt_tokens / cost_completion_tokens
• tricky ids still appear in the per-model breakdown
• a newline in an id no longer splits one record into two lines
• the unset-TRINITY_COST_LEDGER path stays a no-op

Full root suite: 260 passed; ruff clean.

[minion1227]

CI: register_submission failure is not from this PR

The red check is repo-wide infrastructure, not this diff. actions/checkout now refuses to check out fork code under pull_request_target (pwn-request guard), so the job fails at checkout on every fork PR — currently #112, #226, #232, #233, #239, #240 and this one. Because pull_request_target runs the workflow definition from base main, it cannot be fixed from a PR branch. #233 already proposes the fix (dropping the unused checkout step from that job).

All substantive checks pass here: test-router, test-validator, web, and check_submission_body.

Added since opening

A docs/JOURNAL.md entry per AGENTS.md §6, which the original commit omitted — recording that the corruption was silent (ledger_cost_report skips unparseable lines rather than raising, so affected calls simply vanished from the spend total) and why ensure_ascii=False is safe here.