measurement schema 增 run_kind / decode-determinism provenance 欄位（follow-up from #109 verify）

[kiki830621]

Refs #111

Summary

Add two optional provenance fields to the measurement schema (MeasurementRow + SubmissionRow), mirroring the existing hf_revision precedent (optional, nil-default, snake_case CodingKeys, synthesized-Codable free backward-compat):

• run_kind (String?, release-sweep|adhoc|nil) — distinguishes release-sweep full-matrix snapshot rows from ad-hoc submissions so per-version snapshot census completeness is mechanically checkable. New bestasr benchmark --run-kind flag; release-sweep.sh passes release-sweep.
• decode_deterministic (Bool?) — HONEST tri-state: true/false only for backends that consume --decode-deterministic (WhisperKit/whisper.cpp, gated via ModelGrid.backendWhisperKit/WhisperCpp); nil for mlx-audio (silent no-op) and legacy rows. Never claims a value for a backend that ignored the flag.

Re-scoped from Spectra → Plan per user decision (#111 Re-scope comment). The determinism representation upgrade (Bool?+nil vs per-backend enum) is deferred to #118.

bestASR-bench synced separately (direct-commit to bench main, commit e728f1a): validator gains optional-when-present checks (fields kept OUT of REQUIRED so existing 156 rows pass); SUBMISSION_FORMAT.md documents both.

Verification

• swift build clean; 5 new backward-compat tests (legacy decode→nil, round-trip, nil-encodes-absent, packager threads provenance) + full suite 389 green, none skipped.
• bench validator TDD: crafted invalid run_kind/decode_deterministic rejected; 156 legacy rows still OK; valid new-fields rows pass.

Checklist

• Diagnose (re-scoped Spectra→Plan)
• Implement (1 commit, bestASR side; bench synced separately)
• Verify (run /idd-verify measurement schema 增 run_kind / decode-determinism provenance 欄位（follow-up from #109 verify） #111)
• Verify-gated: post-verify PASS = ready to merge → after merge, run /idd-close

---

Generated by /idd-implement on PR path. Do NOT add a GitHub close trailer — IDD discipline requires manual /idd-close after merge.

[kiki830621]

Verify Report — PR #119 (#111)

Engine

manual fan-out (4 lens Agents + sequenced DA, model: opus, file-based output) + Codex (gpt-5.x) — Codex leg incomplete（HTTP 429 usage_limit_reached，同 #112 verify 的 ChatGPT Pro plan limit，reset ~5.4 天）。跨模型盲驗本輪缺席，fail-closed INFO，不當 PASS 計。5 個 Claude reviewer（4 lens + DA）完整。

Aggregate

PASS — 0 blocking, 1 follow-up filed (#120)。

#111（re-scoped Spectra→Plan）verifiably complete。4 lens 全 PASS/APPROVE，DA 認真攻擊後無法推翻核心 PASS，並以 standalone swift snippet 實證兩個最高風險主張。

要求覆蓋率

5 / 5 requirements addressed — FULLY（+ 跨 repo contract exact match）。

Findings（合併後）

#	Severity	Finding	Source	Action
1	PASS	run_kind (String?) + decode_deterministic (Bool?) 加至 MeasurementRow + SubmissionRow，snake_case CodingKeys、threaded 建構點 + denormalization、--run-kind CLI + release-sweep.sh（dry-run + real 兩處 literal）	agents:requirements	✅
2	PASS	backward-compat：synthesized Codable decodeIfPresent（自訂 init 不抑制 init(from:) 合成）→ legacy row decode 成 nil；bench 156 rows 無新欄位仍過（不在 REQUIRED）。DA 以 standalone swift snippet 實證	agents:logic+regression+devils-advocate	✅
3	PASS	nil-encodes-as-ABSENT 非 null（跨 repo load-bearing）：encodeIfPresent 省略 nil key，bench validator 永不會收到 run_kind: null（若收到會拒，因 null ∉ enum）。DA snippet 實證 {"model_id":"m"} 無 key 無 null	agents:logic+devils-advocate	✅
4	PASS	honest decode_deterministic gate 邏輯正確：DA 窮舉 grep 確認 deterministicDecode 恰被 WhisperKitEngine + WhisperCppEngine 讀取（其餘 3 engine 零引用），gate whitelist [backendWhisperKit, backendWhisperCpp] 與 BackendID rawValue byte-identical；mlx-audio 記 nil 不說謊。type-correct（record.backend 為 String）	agents:requirements+logic+security+devils-advocate	✅
5	PASS	無 scope creep：commit 521ce54 恰 7 檔（欄位 + 3 threading + CLI + release-sweep + 1 test）；無 determinism enum（→#118）；IDD tooling state（.gitignore/tree-lock/state）刻意未 commit、亦不入 PR（uncommitted）	agents:regression	✅
6	PASS	security：run_kind/decode_deterministic 未 render 進 leaderboard markdown；CLI 任意字串經 JSONEncoder escape 無法偽造 JSONL row；release-sweep.sh 用 hardcoded literal 非 interpolation；dedupeKey 正確排除 provenance 欄位	agents:security	✅
7	MEDIUM	honest gate 未被單元測試（inline 在 benchmark loop、不跑真 backend 測不到）——「never lie」不變式定義 #111，值得抽 pure helper + 測 4 backend case。gate 邏輯已證正確，缺的是防未來 regression 的護欄	agents:regression	Follow-up → #120
8	LOW	CLI --run-kind 無 client 端驗證（任意字串 verbatim 進 local store，只 bench CI enum gate 擋）——--run-kind typo 產生 submit 時才被拒的 row。5 方收斂	agents:requirements+security+logic+regression+devils-advocate	Follow-up → #120
9	LOW	DA-1：validator 拒 run_kind: null 但收 decode_deterministic: null——run_kind backward-compat 完全靠 encoder 省略 nil（零容忍），對未來 encoder 變更/手寫 row 脆。fix：validator 容忍 run_kind: null 視同 absent	agents:devils-advocate	Follow-up → #120
10	LOW	logic O1：docstring 提 adhoc 但無 code 寫該字面（ad-hoc 走 nil = 同 legacy 不可區分）。doc nuance，census 只在意 release-sweep presence，不影響功能	agents:logic	Noted
11	INFO	cross-model（gpt-5.x）盲驗本輪缺席（Codex 429）——ensemble 降為 5-AI 單模型家族	codex	Process Gap

Scope Check

無超出 #111（re-scoped Plan）範圍。determinism 表示法 enum 正確拆至 #118；hardening 護欄拆至 #120。

Process Gaps

• codex: cross-model leg incomplete — HTTP 429 usage_limit_reached（同 regression-gate.sh：baseline language 欄位未驗證即入 worklist TSV（follow-up from #109 verify） #112，ChatGPT Pro，reset ~5.4 天）。本輪為 5 個 Claude reviewer（4 lens + sequenced DA）。誠實標記、不當 PASS 計；PASS 由 5 個 Claude reviewer 的 empirical 佐證（含 DA 兩個 standalone snippet）支撐。

Follow-up Findings Filed (v2.72.0+ #148)

• harden #111 provenance 欄位：honest-gate 抽 helper+test / CLI --run-kind 驗證 / validator 容忍 run_kind:null #120 — harden measurement schema 增 run_kind / decode-determinism provenance 欄位（follow-up from #109 verify） #111 provenance 欄位（3 項 convergent 護欄）：(1) honest-gate 抽 helper + 單元測 4 case（regression Turn bestASR into a Claude plugin marketplace with an agent-facing plugin (sibling concern from #3) #4，最實質）；(2) CLI --run-kind boundary 驗證（5 方收斂）；(3) bench validator 容忍 run_kind: null（DA-1）。DA 裁決「ship measurement schema 增 run_kind / decode-determinism provenance 欄位（follow-up from #109 verify） #111，track as follow-up」。

紀律聲明

#111 通過驗證，可 merge PR #119 → /idd-close #111。bench 端已 direct-commit push（e728f1a）。#118（determinism 表示法）、#120（provenance 護欄）為非阻擋 follow-up。