fix(cli): assign w{index} ids when loading JSON transcripts

[santhiprakash]

Continuation of #3461, closed in the 2026-09-08 stale-PR sweep with "please reopen with a rebased version if still relevant." This is that rebased version — GitHub blocks author-reopens on PRs closed by a maintainer, so it is filed fresh. Rebased onto current main (d66cd6dcc); the words-json loader on main still defaults missing ids to "", so the fix applies unchanged.

Fixes #3442

What

loadTranscript now assigns id: w{i} on the JSON branches (whisper-cpp, OpenAI, words-json), matching what the srt/vtt branches already do. The words-json branch also stops defaulting missing ids to "" — it repairs them to w{i} instead.

Why

audio/references/transcribe.md documents that word ids (w0, w1, …) are "added during normalization for stable references in caption overrides", but every JSON path dropped them: parseWhisperCpp and parseOpenAI never set id, and the flat words-json branch defaulted it to "", collapsing every word onto the same empty key. All engines funnel through loadTranscript — transcribeAudio rewrites transcript.json from its output — so CLI-produced transcripts ship without ids and per-word caption overrides have nothing to key on (#3442 has the full breakdown, including the SRT round-trip workaround users currently need).

How

Single normalization point: after the existing format-specific parsing in loadTranscript, map id: w.id || \w{i}`over the words.||rather than??so the empty-string ids written by older CLIs into words-json files are repaired on load, not preserved. SRT/VTT branches keep their existing assignment; theWord.id` doc comment already describes exactly this behavior.

Test plan

• packages/cli: bunx vitest run src/whisper/ on the rebased head (c1f87c64b) — 108/108 pass, including updated expectations (whisper-cpp / openai / words-json now carry w0… ids) and the regression test that preserves existing ids (keep-me) while repairing empty ("" → w1) and missing (→ w2) ones. The count includes 4 newer tests from fix(cli): keep phrase-level CJK and Thai transcripts as separate cues #3436 on the rebased base.
• tsc --noEmit in packages/cli — clean on the rebased head.
• Original submission (prior base, see fix(cli): assign w{index} ids when loading JSON transcripts #3461): full packages/cli vitest suite green apart from the same 20 pre-existing environmental failures as clean main, plus oxlint / oxfmt --check green.

This change is AI-assisted; the bug analysis, fix, and tests were verified against the reproduction in #3442.

[santhiprakash]

One CI note so the red X doesn't need investigating: the Tests on windows-latest failure on this head is unrelated to the diff. The only failing test is @hyperframes/engine's ffprobe.test.ts > probeMediaProfile > bounds the process-scoped probe cache, which timed out at its own 5s limit while running 129 fixtures on a cold Windows runner (1 failed / 1706 passed in that suite). This PR only touches packages/cli/src/whisper/normalize.{ts,test.ts}, and the same workflow has been green on 15+ other runs tonight, including three main runs after this one. Re-running failed jobs needs repo-side access, so we're leaving the head as-is rather than pushing a re-trigger commit.