Merge Pipeline, Test Plan
Status: Current Last modified: 2026-10-20 (commit f7e32e22)
This is a map of current test ownership and obligations, not a completion
ledger. Named tests, old implementation-cycle tables and proposed fixtures are
not execution receipts. The removed merge, pipeline and batch CLI
commands are not targets to restore.
See Merge Architecture and Domain Types for the contracts being checked.
Current owners
Paths below are relative to the repository root.
| Boundary | Test owner |
|---|---|
| Structural assembly, ordering, headers, origins and fates | crates/talkbank-transform/tests/integration/transcript_merge_tests.rs |
| Speaker mapping and lexical identification | crates/talkbank-transform/tests/integration/speaker_id_tests.rs |
| Scripted adjudication | crates/talkbank-transform/tests/integration/adjudication_tests.rs |
| Confidence and override representation | Unit modules beside speaker_id/types.rs and speaker_id/override_file.rs |
| Canonical CHAT transform contracts | crates/talkbank-parser-tests/tests/integration/transform_corpus.rs and its declared modules |
| CLI speaker/adjudication/sanity behavior | crates/chatter/tests/integration/ |
| Removed-command help surface | crates/chatter/tests/integration/merge_experimental_help_tests.rs |
The canonical transform harness includes merge_corpus.rs,
gem_merge_corpus.rs, speaker_identification_corpus.rs and
adjudication_corpus.rs. Follow the module declarations for the executable
selection; the existence of a source file alone does not prove it is registered.
Structural merge obligations
Review the assembled model and provenance together:
- Preserve source-relative order, including header and section structure.
- Refuse indeterminate cross-source placement unless the explicit flagged draft policy was requested; retain its review records.
- Preserve untimed speech without fabricating timestamps. Equal recorded starts and absent starts have different meanings.
- Check output origins and every source fate, not just output length.
- Exercise tier stripping as caller policy, including custom and empty lists.
- Exercise language subset admission and conflicting participant metadata.
- Reject invalid assembled output before exposing a reportable result.
- Keep terminal-bullet edit provenance through validation and reporting.
The integration suite contains both success and refusal cases for source-order placement, source anchor reversal, metadata ordering, headers and source fates. A byte-stability assertion for a selected fixture is not a universal promise that model serialization reproduces all original source bytes.
Speaker and wire boundaries
Tests should distinguish lexical no-information, finite and unbounded margins; keep absolute support with the decision. Input admission failure is not a zero-score match. Mapping parsing, mapping application and output validity are different contracts.
The current implementation passes unmapped speakers through. A test for mandatory complete mapping would be a proposed policy change, not a missing assertion to add without adjudication.
Override tests must cover per-speaker rename roles and the current schema rather than the old single-role design. Deterministic local tests of sampling, context preparation and report serialization are not provider executions. Authored context records must be labeled as such; they are not corpus facts.
Choosing evidence
Use spec/reference CHAT for canonical syntax and semantic claims. Boundary and wire tests remain necessary where malformed inputs, external records or deliberately corrupted state cannot be represented by a valid reference transcript. Keep those populations distinguishable.
For a change, select the affected contract deck, inspect the actual assertions, and retain the invocation, candidate identity and complete outcome. Batch related repairs before verification. Do not infer acceptance from a proposed test name, an ignored test, an unregistered module or a historical cycle count.
Documentation-only changes need proportional documentation checks; they do not earn new functional coverage. Follow the current repository workflow rather than the retired one-test-at-a-time implementation schedule.
What these tests cannot establish
Structural correctness and a finite reference population do not establish recording identity, source authority, acoustic correctness, human adjudication, provider quality, corpus readiness, installed-runtime behavior or publication. Each requires its own evidence and authority. No broad corpus run or external provider call is implied by this test map.
This page last changed: 2026-10-20 (commit f7e32e22). The whole book last changed: 2026-10-27 (commit 5bfee2f9).