feat: delegation guidance as per-decision heuristic + escalating tool-cadence hook

[arzafran]

What this does

The delegation reminder fired ~15 times in a single session and got reasoned past nearly every time — repetition without salience, which trains the agent to tune out all hooks. This reworks delegation from a repeated nag into one per-decision question the agent answers before each unit of work, backed by a hook that now fires at most twice per streak: one evidence-based nudge, then one escalation it can't easily skim past.

First (and scoping) PR for #44 — the delegation worked example from the Salience × Memeticity audit. Closes #44; if the same pass should run over the remaining rule set, that's a fresh issue once we've measured adherence on this one.

Summary

• CLAUDE-FULL.md: MUST/SHOULD prose lists → one question (3+ files, 10+ calls, or security-sensitive?) with a routing table, an explicit NO branch, and loop-closing rules (re-ask when scope grows; overriding a YES requires a stated one-line reason)
• src/hooks/tool-cadence.ts (parallelmax branch): now tracks distinct files edited as evidence; one soft nudge per streak (12 calls or 3+ files), then one escalation via PostToolUse block if the streak continues past the reminder — silent after that until an Agent call resets. Was: same paragraph every 12 calls, forever
• config/40-hooks.json: continueOnBlock: true on tool-cadence so the escalation surfaces as a block signal without aborting the turn (already modeled in the zod schema)
• src/hooks/delegation-detector.ts: message compressed to one line, consistent with the heuristic
• docs/hooks-reference.md rows updated (also fixes the stale "N=8" threshold doc)

Test Plan

• bun test — 515 pass, incl. new tests/tool-cadence.test.ts e2e suite (nudge-once, escalate-once-then-silent, 3-file trigger, Agent reset, old-shape state migration)
• bun run typecheck / bun run lint / bun run compose
• Measure adherence over a few live sessions before extending the pass to other rules