Publish text.redact-entities@1.0.0 (PII redaction via NER reuse)

[enricopiovesan]

Summary

Publishes text.redact-entities@1.0.0 -- PII redaction (PER/ORG/LOC/MISC bracketed placeholders) reusing text.detect-entities' (#465) real DistilBERT-NER forward pass directly. Closes #481, the second pick from the post-#473 "Five Engines, What Next" downstream-capability review, after audio.transcribe-speech@1.1.0's VAD gate (#477).

Why a new capability, not a version bump

Unlike the VAD-gate case (a version bump because "transcribe, robustly" was still the same verb), "redact" is a genuinely different action from "detect" for a caller -- different intent, different output shape (redacted_text vs. raw entities). This is a new capability id.

Shape

• capability-src/text-detect-entities gets the same lib/bin split proven on audio-detect-speech-segments for audio.transcribe-speech@1.1.0 — gate Whisper on a Silero VAD speech-presence pre-check #477: engine logic (Model, encode, forward_logits, decode_entities, Entity, plus an always-available fixture_model() for dependents' own tests) moves to a new lib.rs; main.rs becomes a thin _start shim. text.detect-entities@1.0.0's published contract/artifact is untouched -- source-only refactor, all 19 existing tests pass unchanged.
• capability-src/text-redact-entities takes a path dependency on that lib, runs the exact same encode/forward/decode pipeline, then substitutes a bracketed placeholder ([PER]/[ORG]/[LOC]/[MISC]) for each entity span instead of returning raw entity data.
• No new model, no new license/verification burden -- same weight table, same dslim/distilbert-NER (Apache-2.0) license record as text.detect-entities.

Verification

• Verified against the real published model via the compiled wasm32-unknown-unknown --features full-model artifact under wasmtime on real sentences: "Sarah Chen from Acme Corp will follow up with the Berlin team next week." redacts to "[PER] from [ORG][ORG][ORG] will follow up with the [LOC] team next week.", and "Elon Musk founded SpaceX in California." redacts to "[PER][PER] founded [ORG][ORG] in [LOC]." -- both correctly reproducing text.detect-entities' own documented BIO-fragmentation behavior on these exact sentences (verified directly, not assumed to carry over).
• capability-src/text-redact-entities: 10 #[test] cases (placeholder mapping, single/multi-span substitution, no-entities passthrough, defensive out-of-order-span handling, full-pipeline determinism, missing-text-field default); cargo llvm-cov functions=100%, lines=100%, regions=100%.
• capability-src/text-detect-entities: all 19 pre-existing tests still pass post-refactor.

Governing Spec

• 001-registry-foundation
• 014-extraction-compatibility
• 017-persona-registry
• 018-capability-test-coverage

Project Item

https://github.com/orgs/traverse-framework/projects/3 -- #481

Definition of Done

• capability-src/text-redact-entities/ #![no_std]; ABI-clean (imports exactly {fd_read, fd_write, proc_exit}); #[test] coverage functions=100%/lines&regions>=95%.
• capabilities/text/text.redact-entities/1.0.0/contract.json -- full I/O schema, ai.model_backed: true, spec-024 risk.determinism_class: model_derived, spec-023 authoring, provenance citing text.redact-entities@1.0.0 — PII redaction reusing the DistilBERT-NER engine #481.
• text.detect-entities@1.0.0 contract/artifact untouched by the lib/bin source refactor.
• New persona privacy-export-reviewer scaffolded via scripts/scaffold/new-persona.sh (no existing persona fit).
• Signed .wasm, pre_pr_check.sh green, human merge.

Validation

• bash scripts/ci/pre_pr_check.sh <this body> with BASE/HEAD wired as CI wires them.

🤖 Generated with Claude Code