docs(api): api-review daily audit 2026-09-01 — localize english text in ZH specs (336 ops, zero surface change)

[flashduty]

api-review daily audit — 2026-09-01 (UTC)

Mode: --mode generate --scope all --auto (equivalent to /api-review --mode generate --scope all --auto)
Registry source: fc-pgy/logic/api/api_test.go @ a21dfb3 (2026-09-01 13:26 +0800), 926 rows parsed
Docs baseline: main @ e50f58f

Operation surface: zero drift

Public-scope filter (Auth == "all", path not /event/push/*, non-hidden module) yields 336 operations; the committed specs contain exactly the same 336 (method, path) pairs. Set equality was verified both directions.

Module	Ops in spec	added	updated (schema/shape)	removed
on-call	192	0	0	0
monitors	39	0	0	0
rum	36	0	0	0
platform	27	0	0	0
safari	42	0	0	0
total	336	0	0	0

The only registry change since the last run (2026-08-31) is PR #879 feat/huawei-aom, which adds 4 Huawei cloud push rows (/event/push/alert/huaweiyun/aom, -apm, /ces, /lts). All four are auth: integration, so they are out of public API scope by design and correctly absent from the specs — no new public endpoint to document.

Handler-side scan across fc-event, fc-oncall, fc-pgy, fc-rum, fc-statuspage, fc-datasource, go-pkg for commits since 2026-08-31T08:12Z found no change to any public API request/response shape: the fc-event work is AI SRE IM-card/feedback rendering (logic/ai_sre/*), fc-oncall/fc-rum/fc-statuspage/go-pkg had none in controller or struct paths, and the fc-datasource change is PlugTeams attachment download.

What this PR changes: bilingual-contract text drift (146 lines, text-only)

No structural change — verified by full-tree deep compare: paths keys, components.schemas keys, operationId list and tags-entry shape are all byte-identical to HEAD. Only four leaf text fields moved, and only these:

Count	Change	Class
104	200-response description: "Success" → "成功" (in the .zh files)	ZH file carrying English text
30	op tags[0]: "On-call/Incidents" → "On-call/故障管理"	English tag in ZH file, and undeclared in the ZH tags[] list
6	tags[] name + 4 op tags[0]: "RUM/Session replay" → "RUM/RUM 会话回放"	English tag in ZH file, diverging from mapping.yaml
2	/account/info 200 description: "OK" → "Success" (EN) / "成功" (ZH)	one-off against the 330/331-file convention
2	/rum/issue/export 200 description → "成功。CSV 附件，非 JSON 信封。"	English text in ZH file; wording copied from the sibling ZH export ops

Files touched: on-call.openapi.zh.json (26), safari.openapi.zh.json (41), rum.openapi.zh.json (4), platform.openapi.{en,zh}.json (1 each), plus the two consolidated copies kept in sync (openapi.zh.json 72, openapi.en.json 1). Total: 7 files, 146 insertions / 146 deletions.

Rationale for the ZH tag renames: mapping.yaml is the authority for tag labels (<parent_zh>/<tag_zh>). on-call/incident declares tag_zh: 故障管理, and 40 of its 55 ZH ops already carried On-call/故障管理 — the other 15 (/incident/feed, /incident/comment, comment-type/*, work-item/*) carried the English label, which Mintlify renders as a separate English sidebar group. rum/session-replay declares tag_zh: RUM 会话回放; the spec carried the English form. RUM/RUM Sourcemap was left alone — it looks untranslated but is exactly mapping.yaml's declared tag_zh ("RUM Sourcemap").

Not changed (deliberate exemptions, all evidence-checked)

• docs.json and {en,zh}/openapi/api-catalog.mdx: untouched. Zero operations were added or removed, and every one of the 336 spec paths is already present in the nav allow-list and in both catalogs (verified programmatically) — the 0-endpoint-change case where 防错铁律 3 forbids nav edits.
• openapi.legacy.zh.json: untouched (read-only reference).
• 16 "missing 200 example" + 4 "missing request example" audit hits are false positives: those operations are non-JSON by design (text/csv, application/octet-stream, application/x-ndjson responses; multipart/form-data requests). The repo's own scripts/lint_openapi.py codifies exactly this exemption and passes.
• 30 "epoch wording" audit hits are false positives. After disambiguating durations from epochs, the only remaining flags are genuine durations that correctly omit epoch wording: ScheduleNotify.advance_in_time (lead time in seconds), ServiceMap*.max_age_ms / report_interval_ms (data age / reporting interval). Zero-milli violations among real ms-epoch fields.

Verification run in this session

• python3 scripts/lint_openapi.py → OK: 12 spec files, no violations
• All 12 output JSON files parse (json.load), with path/schema counts printed per file
• Full-tree leaf compare vs git show HEAD:<path> → 146 diffs, every one in the allow-listed table above; zero unexplained
• paths / schemas / operationId / tags-shape identical to HEAD (no structural or ordering churn)
• split ↔ consolidated: 336/336 ops, 0 content-differing, both languages
• registry ↔ spec set equality: 336 == 336, new=[], stale=[]
• Baseline fidelity: all 12 files are byte-identical under a json.dumps(indent=2, ensure_ascii=False) + trailing newline round-trip, so no reformatting could leak into the diff. Every edit was applied as an assertion-guarded line replacement (expected count per file per class had to match exactly, or nothing was written).

findings.unresolved / environment notes

• runbooks/api-review-apply-patches.py and runbooks/api-review-daily.md are still absent from the team knowledge pack (same as the 2026-08-28 note), so 防错铁律 6's "patch first" step could not run. Baseline fidelity was instead guaranteed by the round-trip proof + assertion-guarded line edits described above. No full regeneration was attempted with the unpatched generator.
• mapping.yaml coverage gap (skill-side, not fixed here): 13 paths that are auth: all in the registry and already documented in the specs are claimed by no path_prefixes entry, so a --scope all analyze run would drop them: monitors — /monit/query/data, /monit/query/diagnose, /monit/targets, /monit/tools/catalog, /monit/tools/invoke, /monit/servicemap/{status,summary,topology,fleet,fleet/summary}; rum — /rum/data/query, /rum/field/list, /rum/resource/info. Recommend adding these prefixes to mapping.yaml (or to the missing patch script) before the next generate, otherwise a regeneration would delete 13 live endpoints.
• monit-webapi / monit-edge are not on GitHub (flashcatcloud org), so the monitors module could not be re-extracted from handler source this round and carries forward unchanged from HEAD apart from nothing (0 lines touched).
• No constructed examples were added this round (no new operations), so there is no synthetic-example caveat.