v0.5.0 — Offline-First Build Pipeline Release

v0.5 Draft (2026-04-26)

Status: RFC. Introduces the v0.5 offline-first build pipelines (ASR / text /
vectorization / moderation), a derived-asset provenance schema, and the
single-entrypoint pipeline CLI. No breaking changes to v0.4 manifests; the new
pipelines write everything under derived/<name>/ so existing records are
untouched until a pipeline is explicitly run against them.

Added

• pipelines/ directory with the v0.5 pipeline contract:
  • pipelines/__init__.py — PipelineSpec registry + dispatcher.
  • pipelines/_descriptor.py — shared DescriptorBuilder that emits
    <output>.descriptor.json validated against
    schemas/derived-asset.schema.json.
  • pipelines/asr/ — dummy (deterministic, no model) and faster-whisper
    (lazy-imported, opt-in) backends.
  • pipelines/text/ — NFKC normalisation + conservative redaction
    (priority order: URLs with embedded credentials, emails, CN ID
    cards, CN mobile phones, IPv4 addresses, credit-card-like 13–19
    digit runs, generic phone numbers). Replacements use stable
    category placeholders (<EMAIL>, <PHONE_CN>, <ID_CN>, <IPV4>,
    <CARD>, <PHONE>, <URL_WITH_CREDENTIALS>). redactions.json
    sidecar carries rule_name + start/end + replacement only and is
    auditable without re-leaking matched substrings.
  • pipelines/vectorization/ — paragraph-aware chunking with absolute char
    offsets, hash (deterministic 64-D) and sentence-transformers backends,
    optional Qdrant push (backend and model_id stored as separate
    payload keys so downstream filters work without ambiguity).
  • pipelines/moderation/ — deterministic regex/wordlist policy with
    severity-based outcome aggregation (pass | flag | block). Built-in
    v0.5 policy + --policy-file for JSON/YAML overrides. Flags carry
    rule + span only, never the matched substring.
• tools/run_pipeline.py — single CLI entrypoint (python tools/run_pipeline.py <name> --record path/to/record …) shared by every pipeline.
• tools/validate_pipelines.py — static guard: enforces the
  derived/<spec.name>/ output-prefix invariant and refuses any module that
  imports a hosted-API client (openai, anthropic, google.generativeai,
  cohere, aliyun_sdk_bailian, …). This is what turns "offline-first" into
  machine-checked policy.
• tools/test_pipelines.py — umbrella test driver. Runs the four
  per-pipeline test scripts as subprocesses so an import failure in one
  pipeline cannot mask test results in another.
• tools/test_asr_demo.py — end-to-end test for examples/asr-demo.
• schemas/derived-asset.schema.json — provenance descriptor schema
  (schema_version / derived_id / record_id / top-level pipeline +
  pipeline_version / actor_role / inputs.{source_pointers,inputs_hash}
  / output.{path,outputs_hash} / optional model.{id,version?,source?, online_api_used: false} (required when pipeline is asr or
  vectorization) / optional moderation_outcome).
• examples/asr-demo/ — self-contained fixture record. run_demo.sh
  regenerates a deterministic placeholder WAV (DLRS is pointer-first so
  audio is never committed) and walks all four pipelines end-to-end with
  no model download.
• docs/PIPELINE_GUIDE.md — companion to the example. Covers the contract,
  the descriptor, every pipeline's CLI, authoring guide, and what v0.5
  deliberately is not.
• .github/workflows/validate.yml: dedicated pipelines job parallel to
  validate, matrix over Python 3.11 and 3.12.

Changed

• tools/batch_validate.py: collapsed the four per-pipeline tests into a
  single pipelines step delegating to tools/test_pipelines.py, then
  added asr_demo for the end-to-end fixture. Local report:
  11/11 passed.
• docs/GAP_ANALYSIS.md and docs/IMPLEMENTATION_STATUS.md rewritten to
  reflect v0.5 (overall completion ~83%).
• ROADMAP.md: v0.5 marked as released, with the Closes #N-per-PR
  governance rule appended to the v0.5 section so future major versions
  inherit it.

Closes

#28 (epic), #29, #30, #31, #32, #33, #34, #35, #36, #37, #38.

---

Sub-issues closed by this epic

#28 (epic), #29, #30, #31, #32, #33, #34, #35, #36, #37, #38.

Validation

tools/batch_validate.py --report-dir reports -> 11/11 passed.

Compare

v0.4.0...v0.5.0