transcribe phase 0: correct the diff target before implementing against it

[fyang0507]

Context

Decomposes #12, phase 0 of
TRANSCRIBE_IMPLEMENTATION_PLAN.md.

TRANSCRIBE_HAPPY_PATH.md is the diff target for every phase after
this one, and tests/test_spec_docs.py, tests/test_environments.py, and
tests/test_shipped_commands_match_the_document.py enforce it. An evidence pass over the
runners and recorded artifacts found six things in the specs that the artifacts refute. They
land first, because implementing against a wrong target produces code that matches the wrong
number.

Corrections

Each is a value or a claim an artifact or a pinned source refutes.

• bytes: 2463307541 → 2467859030, five occurrences (CONTRACT §1.1 and §1.2,
  HAPPY_PATH §1.2). manifest.json carries the measured figure read from the Hub at the pinned
  revision, and tests/test_environments.py:205 retires this exact number as one of the
  illustrative ones the specs used.
• HAPPY_PATH §1.1 word_timestamps note says the aligner is "in the torch environment,
  so this request spans three". It moved to mlx beside the ASR and adds no runtime — see
  ENVIRONMENTS.md and CONTRACT §1.4, which already say so.
• HAPPY_PATH §2.1 execution.note says "both model stages share the torch environment"
  one line below environments_spanned: ["mlx", "torch-vibevoice"].
• HAPPY_PATH §3.1 residency and peaks. run_firered.py:395-426 loads VAD, LID, ASR and
  Punc, assembles them onto one system object (:435), then calls system.process once
  (:463) — so residency: "one_model_stage_at_a_time" and "four model stages in one
  environment, none resident together" are false on the only stack with four stages. Measured on
  the same 139.284 s clip: 9.16 GiB peak RSS with LID off, 12.26 GiB with LID on; the
  30-minute channel is 9.12 GiB with LID off, so the peak is dominated by weights rather than
  duration. §3.1 publishes a LID-on run at 6.43 GiB, below the measured LID-off floor,
  because it was built as a maximum of isolated per-stage peaks. Correct residency to the
  granularity that holds — no two environments resident at once, FireRed's models co-resident
  inside torch-firered — and replace the illustrative peaks with the measured pair.
  Artifacts: firered_lidoff_batch4_multispeaker_codeswitch_20260815.json,
  firered_lidon_batch4_multispeaker_codeswitch_20260815.json,
  firered_lidoff_batch4_spice30m_participant.json.
• unit_count_known_at_plan_time is described in CONTRACT §1.1 prose and emitted
  nowhere, and VOCABULARY's "unit" entry says the count is "reported absent" where the payload
  prints null. Ruling: null in a fixed-shape structural block, because a processing block
  whose key set varies by stack is worse than a null with a note saying why. VOCABULARY's
  sentence changes.
• Attribute the diarization figures to the configuration that produced them. The
  95.42 % participant-interval F1 and the 33.56 % MER both trace to a diarization run with
  --num-speakers 2 and --overlapping-segments, and that run shaped the turn set the MER was
  scored against (589 raw segments → 195 accepted turns, 9 overlap abstentions, 33 short turns,
  54 raw-fragment-only spans). The shipped configuration supplies neither flag by default, so
  the diarization and overlapped_speech notes say which configuration the figures describe
  and that the shipped one is unmeasured.

Additions

Names and rows the implementation needs that no document carries yet.

• abstentions[].reason: overlap (more than one speaker active), short_turn (an
  accepted turn below the minimum), raw_fragment (a span whose only activity was a sub-250 ms
  diarizer fragment). Three causes the recorded runner distinguishes; the budget-unprocessed
  turns it files beside them are coverage, not abstention — nothing was declined, the work
  was not reached.
• The three turn-threshold values VOCABULARY names without publishing:
  raw_fragment_min_ms 250, accepted_turn_min_ms 500, same_label_merge_max_ms 300
  (model_tests/benchmark/run_turn_attributed_mlx_asr.py:78-84).
• failure_recovery.partial_results gains prefix_only; none is retired to VOCABULARY's
  retired words. vibevoice was its only holder and a truncation leaves a usable prefix, so
  none becomes a member nothing reaches — the satisfaction: unavailable mistake. The
  mechanism is in the vibevoice phase issue.
• complete in a result document, always present, with coverage only when it is
  false. The one absence that would be dangerous rather than meaningful: a saved document
  outlives its exit code and export branches on it when merging.
• One error code, option_value_unsupported (exit 2; field, provided, allowed,
  did_you_mean when a near value exists, fix) — a stack that takes the option but not that
  value. option_unsupported_on_stack keeps its exact field list for a stack that takes no such
  option. Amending one row to carry conditional fields was considered and rejected: the
  three-way capability split exists so two different failures are not rendered alike.
• The 30 accepted --language names, cited to config.json:support_languages at both
  pinned Qwen revisions, and FireRedLID's 115-label vocabulary cited to its dict.txt.
• roles.aligner.config.language_rule, and the Cantonese trap as its stated reason:
  qwen3_forced_aligner.py:256-266 branches on japanese, korean, chinese, and everything
  else to tokenize_space_lang, while its checkpoint declares Cantonese as supported. Passing
  Cantonese through would space-split CJK into one word per segment.

Code

• A backend→(package, role) map in manifest.json.
  environments.packages_for() keys roles.values() on package ids
  (src/audio_cli/environments/__init__.py:141), so firered-vad, firered-lid,
  firered-asr2-aed, and firered-punc — four backends of the single firered-asr2s package —
  raise ManifestError. That is package data and belongs beside the rest of it.

Acceptance

• uv run --extra dev pytest green with every edit in place.
• packages_for("firered", {…four backends…}) returns [firered-asr2s], and still raises
  on a backend that does not fill the role it is mapped to.
• No document states a residency claim the recorded runs contradict, and no spec figure
  equals a value tests/test_environments.py retires.