Add judgment-layer mistake entries to AGENTS.md

[nateshpp]

What this PR does

Adds three entries to AGENTS.md's mistake catalog (section 4) and one trigger to its
escalation list (section 6). All three name a failure mode no deterministic gate can
catch: silently picking one reading of an ambiguous goal, building past the stated
need, and drifting into unrelated code inside an otherwise in-scope file. Today these
get caught only in review, if at all. This gives reviewers, human or agent, a citable
rule instead of ad hoc pushback. Tracked as WI-054.

Details

• Entry 19, silent interpretation. A goal with two or more materially different
  readings gets named and escalated (in session, or as a DECISIONS.md Open entry with
  options and a recommendation), never silently resolved one way.
• Entry 20, speculative scope. No feature, config control, or abstraction beyond the
  stated need. Every config control already costs four synced representations plus a
  changelog entry (section 5); this makes the discipline explicit and citable.
• Entry 21, within-file scope drift. Complements entry 9: entry 9 polices which
  files change (the edit-set gate); entry 21 polices the lines inside a file the item
  already allows, where no gate can see. Unrelated dead code gets flagged in review, not
  deleted.
• Section 6, item 10. New escalation trigger: a goal that supports more than one
  materially different reading routes to the owner with the readings and a
  recommendation, reusing the options format DECISIONS.md Open entries already carry.
• Documentation only. No script, schema, or config control changes; schema_version
  stays at 1.

Governance checklist

• npm run verify passes locally (drift guard, style check, tests, AgentProof)
• No em dashes in any tracked file
• Prose states conclusions directly
• No AI co-author lines in commit messages
• N/A: no ratchet logic changed
• N/A: no new AgentProof scenario
• N/A: no config schema changed; CHANGELOG entry added regardless
• N/A: prompts/ untouched, no bundle rebuild needed

Test evidence

# tests 914
# suites 17
# pass 914
# fail 0

Score: 25/25 normative | 10/10 extended (35/35 total)
Level: HARDENED