Format audit against Cassini main

Compared against fetched gocassini origin/main, commit e4856e9fac32c447d6c5e023681433319d7ea3c8. The local Cassini checkout was behind at 7774ed1; it was not used as the comparison target. The matching implementation is Cassini PR 276, merged to main as 4c7f8bd.

The main simplification is to stop classifying the origin of word-timed text. The words, their body format, the selected default and optional processing provenance already describe what a reader needs. Requiring a label buys only extra validation and special cases.

Removed from the current contract#

Feature	Evidence on the pinned main	Change
Word-transcript role and sourceTranscriptId	Go and the viewer validate an origin enum. The normal writer hardcodes raw-asr; authored-text labels are fixture metadata. No production correction or translation workflow was found.	Word entries need id, format, payloadRef; defaults and array order select them. Ignore old origin labels and links.
readable-cleanup and provenance.readableCleanup	manifest.go calls the role withdrawn; the cleanup producer has been deleted. Go explicitly skips older entries.	Remove the current cleanup slot and provenance definition. Keep old entries skippable.
chapters	Go has a struct and JSON round-trip plumbing, but no portable writer populates it. The viewer's optional chapters.vtt sidecar is separate.	Remove the manifest field and chapter schema.
meeting.summary	The Go field's comment says it is reserved, left empty, and has no requested meaning. The writer stores real summaries in summary.md.	Remove this unused second place for summary content.
meeting.language	The production manifest builder never sets it. Transcript language and processing-step language already exist.	Remove the meeting-level declaration; keep transcript and processing language.
meeting.roomName / CASSINI_ROOM_NAME	Already removed from the Go meeting struct and writer; the current room name belongs in the catalog.	Remove the current declaration and mirror tag. Old members follow the unknown-member rule.
payloads / CASSINI_PL_	A future reservation in this spec, with no writer or reader implementation found.	Remove the reservation.

Implementation evidence:

• Go manifest types and validation.
• Transcript encoding, validation and defaults.
• Production manifest construction.
• Production transcript assembly.
• Viewer validation and selection.

The current source and checked-in proposals were searched for writers and consumers. An unused type or an old design's hypothetical use case was not counted as an implemented feature. This cannot establish plans outside the repo.

Kept because Cassini uses it#

Feature	Current use
Multiple transcripts, ids and default selection	The producer packs additional model passes; the viewer switches among them and gets labels from ids and provenance.
readableTranscripts, role: "display", sourceTranscriptId	The producer packs native transcript.display.v1 blocks. The viewer needs the source link to pick the matching display when switching transcripts.
provenance.speechToText and displayTranscript	Per-pass model descriptions and display processing records.
Attribution fields and provenance.wordTimings	The pipeline records crosstalk evidence; timing provenance controls whether a consumer repairs word ends. Removing it changes behavior.
Top-level summary, attachments and summary provenance	Real summary.md content and its generation metadata are packed and read.
Room id, job id, attempt number, recording and processing times	Actual operator output and artifact identity.
Chunk descriptors, independent bodies and audio integrity	The readers validate these and use the audio digest to bind metadata to the recording.
x private metadata	The site's synthetic demo records how its audio and timings were made here.

Some duplication could be removed in a larger wire revision: audio versus integrity, constant profile/codec/rate tags, MIME and encoding descriptors, and word counts. Current code uses or validates these. They are not removed by this pass. The ten reserved transcript ids are also enforced by the Go writer and validator; their stated namespace-collision rationale is dubious because transcript chunks already live under CASSINI_TX_, but removing that restriction can be handled separately.

Compatibility and Cassini follow-up#

The new readers still accept existing word transcripts with origin labels. Historical drafts and the original conformance vectors are kept as such. The new producer and demos omit the label. The audio and payload transport retain their existing format ids.

The audited main rejected a word entry with no role. The Cassini follow-up implements these changes:

1. Go and browser validators ignore word origin labels and source links, including non-string legacy values. Display source and payload validation remain.
2. Build-artifact readers ignore legacy word roles; new build artifacts and portable output omit them. The internal encoder distinguishes only words from display documents. Existing raw-asr ids stay stable and opaque.
3. The static exporter follows the same rules as the viewer. Unsupported readable roles, including withdrawn cleanup, are skipped without loading their bodies. Displays are selected by source, then default, then array order.
4. The vendored schema and format guide match the public contract. Unused Go chapter and meeting-hint fields are removed.

Readers continue accepting older files. Installed Cassini versions that predate this follow-up need an update to open newly produced files without role; merging source changes does not upgrade an installed ExApp.

The display body correction is separate from origin labels: the old spec incorrectly called every body cassini.words.v1, even though Cassini writes and reads native display blocks. The current spec now reflects that behavior.

Validation#

The JavaScript reader and both Python extractor profiles have zero failures across 26 vectors. The four new vectors cover omitted/unknown origin labels, second-default selection, and a native display beside an unavailable withdrawn cleanup entry. Schemas validate the new files. The original 22 vectors retain legacy metadata for compatibility coverage.

Cassini validation: recorder go test -race ./... and go vet ./...; 428 viewer tests and 99 app tests; standalone and embedded builds. Tests cover legacy build-artifact input through packing, non-string ignored metadata, display source validation and selection, transcript switching and crosstalk mapping.

The updated Go binary passes all four new vectors. Across the full suite it has 18 passes, 7 advisory warnings and one failure: 011, where it rejects a tag/manifest audio-digest disagreement instead of using the manifest. That failure was reproduced on the pinned starting main and predates this change. The old main also rejected 023–025 for their missing/unfamiliar roles; its viewer showed the same errors. Those origin-label failures are resolved by the follow-up.

The site passes type checking and a production build, with existing CSS and accessibility warnings. All three tracked demos validate; their audio packets and word bodies match the previous files exactly.

An independent agy --dangerously-skip-permissions review identified the static exporter, legacy build-input compatibility, and missing display/cleanup coverage. Those paths are included in the implementation and regression checks above.

The final review confirmed those fixes and explicitly agreed to merge after CI, with no blocking defects. The review record is attached to Cassini PR 276.

All required Cassini CI checks passed before merge, including the installed ExApp Talk recording check. That check needed one retry after a Go module proxy download failed before the stack started; the unchanged commit passed.