feat(challenge): add Detection Noise vocabulary to normative-vocabulary

[klappy]

Move challenge detection stop words from oddkit code into canon governance

What

Adds a new ## Detection Noise section to odd/challenge/normative-vocabulary.md as a code block of common filler words. Updates the blockquote, Summary, and Notes to acknowledge the article's now-dual scope: signal in retrieved canon quotes (existing) plus noise in user input matched against per-type detection text (new).

Why

The oddkit worker currently hardcodes a CHALLENGE_STOP_WORDS Set in workers/src/orchestrate.ts — a Vodka Architecture violation in a refactor (oddkit#100) that was explicitly about removing such violations from source code. The hardcoded constant carries a domain opinion ("modals are signal, articles are filler in challenge detection") that belongs in canon, not in worker source.

Caught in PR #100 review by Klappy. The gauntlet didn't surface it — the category "is this the right architectural shape" requires a different lens than the current tools provide.

Scope decision

Single article, two surfaces — chose Option A from the discussion. Pros: one fetcher pattern, drift-free domain vocabulary, atomic edits when extending or pruning to a new domain. The two surfaces are two roles of the same domain opinion ("what counts as content vs filler in this domain").

Modal verbs are deliberately absent from the filler list

must, should, shall, may, might, can, could, will, would, not, no, never, always, do, does, did, have, has, had are all NOT in the Detection Noise list. They are the load-bearing trigger words for the strong-claim, proposal, and assumption challenge types. Filtering them would silently break those type detections — exactly the bug the BM25 pivot in oddkit#100 caught.

Companion PR

oddkit#101 (will follow): drop the hardcoded CHALLENGE_STOP_WORDS constant, extend fetchNormativeVocabulary to extract the new section into a Set, return it on the vocab object, consume it in discoverChallengeTypes when building the per-type BM25 index. Backward-compatible — empty Set when section is absent (server falls back to no filter, IDF only).

Verification

• AI voice clichés audit on new prose: clean
• Summary section preserved (Writing Canon tier 2 requirement)
• Frontmatter governs field broadened to reflect dual scope
• Frontmatter date bumped to 2026-04-17

Refs

• Caught in: feat(challenge): governance-driven runChallengeAction (E0008) oddkit#100 review feedback
• Will unblock: oddkit follow-up PR (drop hardcoded constant, read from this section)

---

Note

Low Risk
Documentation-only canon governance changes; no runtime code changes in this PR, with low risk aside from downstream tooling interpreting the new ## Detection Noise section incorrectly.

Overview
Updates odd/challenge/normative-vocabulary.md to explicitly cover two detection surfaces: signal words/phrases in retrieved canon quotes (tension detection) and a new ## Detection Noise stop-word list for filtering user input before BM25 scoring.

Adds the noise vocabulary section (code block of common filler words) and revises frontmatter, summary, and notes to describe how the server should parse/apply both vocabularies, including fallback behavior when the new section is missing/empty.

^{Reviewed by Cursor Bugbot for commit 7c0e65f. Bugbot is set up for automated code reviews on this repo. Configure here.}