handoff: P1.3.4 encode canon-parity refactor (D5 + D9) — closes the sweep

[klappy]

Forward handoff for the next session: P1.3.4 — migrate oddkit_encode's trigger-word classifier from regex alternation to stemmed set intersection (D5) and remove the module-level cachedEncodingTypes cache (D9). This closes the canon-parity sweep.

Why this is next

Encode is the only tool in the sweep whose matcher still uses literal regex alternation:

new RegExp("\\b(word1|word2|...)\\b", "i")

The trigger vocabulary itself is already governance-driven — runEncodeAction's parse loop reads ## Trigger Words from each odd/encoding-types/*.md at runtime. What's NOT governance-driven and NOT stemmed is the matcher. Input "deciding" does not match canon vocab decided because \b(decided|decision|chose|...) is literal word-boundary alternation. Input "realizing" does not match realized. Same bug class challenge had pre-0.21.0.

Gate (P1.3.2) and challenge (P1.3.3) both closed their vodka anti-pattern remnants. Encode is the last holdout.

What's in the handoff

• Bootstrap-first instructions. First turn of the next session: oddkit_time, oddkit_get on klappy://canon/bootstrap/model-operating-contract, read the "Before Shipping Code" section, follow the pointers to release-validation-gate and contract-governs-handoff-drift.
• Current state: prod at 0.21.1, main at d17bc0c, canon at bc6bfb1/ee9aee4. No pending canon-first prereqs.
• Scope — exactly two items: D5 stemmed matcher for EncodingTypeDef.triggerRegex → stemmedTokens, and D9 removal of cachedEncodingTypes / siblings at L506–508.
• Load-bearing design callout. The no break comment at orchestrate.ts L1161 is canon-by-convention — parseUnstructuredInput emits one artifact per matching type intentionally ("We must never deploy without tests" is both Decision AND Constraint). The refactor preserves both the first-match path (L1123, batch-untagged) and the multi-match path (L1165, parseUnstructuredInput) with their existing contracts.
• Stop-word caveat explicit. Pass new Set() to tokenize() on both sides. Canon trigger vocab includes going with, committed to, turns out, next step, blocked by, must not, haven't decided — all contain stop-words. P1.3.3 Bug feat(website): implement public website for PRD v1.0 #1 is the concrete precedent.
• Stemmer gemination caveat. From P1.3.2. Audit each canon trigger vocab for geminating inflections before shipping.
• Exact line numbers for every touch point.
• Smoke expansion — 5 new regression assertions (inflection match, stop-word survival, multi-type preservation, first-match preservation, cache-removal structural check). Target smoke count: 176–177 pass.
• CHANGELOG sketch for 0.22.0.
• Validation plan — single path, binding. Deliberately no "Option A vs. Option B" section. The P1.3.3 handoff had one and it produced the P1.3.3 incident. This handoff makes the release-validation-gate path mechanical: wait for Bugbot on both feature + promotion PRs, dispatch Sonnet 4.6 validator (5-corroboration adapted for encode) before promotion, fold findings into closeout ledger, then merge.
• Attestation mapping table — each constraint tied to the specific gate that satisfies it.
• Carry-forward O-opens updated from P1.3.3, including P11 (mechanical gate enforcement of release-validation-gate) flagged as the natural post-sweep follow-up.
• Thin prompt at the bottom for fast next-session context load.

Provenance notes

The handoff is explicit about its own authoring context: it's written immediately after P1.3.3 — the session that demonstrated why release-validation-gate needed to exist. The validation plan is single-path on purpose. If the next session reads anything in this handoff that conflicts with canon, canon wins per contract-governs-handoff-drift.

Writing-canon gate satisfied

• Blockquote with compressed argument ✓
• Summary sections (Bootstrap + Current state) ✓
• Descriptive headers (each section header tells the story when scanned) ✓
• Frontmatter well-formed, derives_from references real ledgers + canon ✓
• Cross-references to related canon throughout ✓

What closing the sweep unlocks

Before sweep	After sweep (post-0.22.0)
Encode uses regex alternation	Stemmed set intersection
Challenge uses regex per check	Stemmed set intersection + 4 structural flags (shipped 0.21.1)
Gate uses hardcoded word-boundary regex cascade	BM25 for transitions, set intersection for prereqs (shipped 0.20.0)
3 module-level derivation caches	0 module-level derivation caches (fetch layer caches still)
No canon governing the release pipeline	release-validation-gate + contract-governs-handoff-drift (shipped ee9aee4)

Next move after the sweep closes: P11 — mechanical enforcement of release-validation-gate in oddkit_gate at execution → completion transitions. That's a capability add, not a parity ship — candidate for an epoch transition (E0008.4 or E0009), not P1.3.5.

Refs

• Predecessor ledger: klappy://odd/ledger/2026-04-20-p1-3-3-challenge-canon-parity-landed
• Release validation gate (binding): klappy://canon/constraints/release-validation-gate
• Contract governs handoff drift (binding): klappy://canon/principles/contract-governs-handoff-drift
• Cache fetches and parses: klappy://canon/principles/cache-fetches-and-parses
• Vodka architecture: klappy://canon/principles/vodka-architecture
• Managed agents skill (Rule 2 dispatch): /mnt/skills/user/managed-agents/SKILL.md

---

Note

Low Risk
Adds a new handoff markdown document only; no production code or configuration changes are included.

Overview
Adds a new odd/handoffs/2026-04-20-p1-3-4-encode-canon-parity.md handoff describing the next session’s planned oddkit_encode canon-parity refactor: switch the trigger-word matcher from regex alternation to stemmed token set intersection (while preserving first-match vs multi-match semantics) and remove the in-process cachedEncodingTypes cache per cache-fetches-and-parses.

Includes an updated, single-path validation plan aligned to release-validation-gate, plus concrete smoke-test assertions and a 0.22.0 changelog sketch to guide the implementation and release.

^{Reviewed by Cursor Bugbot for commit c090381. Bugbot is set up for automated code reviews on this repo. Configure here.}

[klappy]

Release Validation Gate — Disposition

Rule 1 (active reviews): Cursor Bugbot completed/success on c0903817, no findings. All checks completed. Rule 1 satisfied.

Rule 2 (independent validator for load-bearing surface): This PR is doc-only — adds odd/handoffs/2026-04-20-p1-3-4-encode-canon-parity.md, zero workers/src/ changes, zero canon changes. Per klappy://canon/constraints/release-validation-gate the Rule 2 trigger is "any change to load-bearing surface" and the not-triggered list explicitly includes "Documentation-only PRs." Rule 2 not triggered; no validator dispatch required.

Rule 3 (canon outranks session artifacts): This PR's content is itself a session artifact. Its validation plan section is written in single-path form specifically because the P1.3.3 handoff's optionality produced the P1.3.3 incident. If the next session reads anything in this handoff that contradicts canon, canon wins per contract-governs-handoff-drift. The handoff's Provenance section names this explicitly.

Merging.