fix(hooks): run the deterministic Truth Gate on Claude too

[CodeWithJuber]

Summary

Two changes, both small and both evidence-backed.

1. The Claude adapter never invoked the deterministic gate

hooks/codex.json wires the Stop event to a deterministic command:

{ "type": "command", "command": "sh ${PLUGIN_ROOT}/hooks/truth_gate.sh" }

hooks/hooks.json — the Claude adapter — wires it to a probabilistic LLM check:

{ "type": "prompt", "timeout": 20, "prompt": "You are Truth Gate, ..." }

But hooks/truth_gate.sh line 4 is:

ROOT="${PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-}}"

The deterministic gate already supports Claude Code — I confirmed by execution that
CLAUDE_PLUGIN_ROOT behaves identically to PLUGIN_ROOT (both reach the fallback ladder and block
correctly on a completion-claim-plus-unfinished-marker payload). The Claude adapter simply never
called it.

This looks like a migration oversight, not a design choice. hooks/hooks.json has one commit
(eb2b88dc, v2.0.0). hooks/codex.json has that plus 2c72120d (v3.0.0) — the release that shipped
the Rust kernel, added truth_gate.sh, and rewired Codex to prefer the deterministic binary. That
commit touched 67 files; hooks.json was not among them.

The fix is additive. The command hook runs before the existing prompt hook, whose text is
preserved byte-identical. Claude gets the deterministic floor and the semantic pass — the prompt
hook catches "deliverable missing / uncertainty hidden" cases a script cannot, and the script
guarantees a floor the prompt cannot.

2. docs/EVIDENCE.md staleness pass (rule 5)

Applying the doc's own maintenance rule. Nothing existing was rewritten — this only appends:

• Denominator footnote. The existing 5.2% / 21.7% figures are code-sample level (576,000 samples).
  The same study also reports package-mention level: 2.23M mentions, 440,445 hallucinated (19.7%),
  205,474 unique names, plus a separate rerun experiment where 43% recurred in all ten reruns. Both are
  correct; readers hitting the other number elsewhere now know why.
• 2026 replication. Five frontier models, rates compressed to 4.62%–6.10% (arXiv:2605.17062) —
  explicitly flagged single-author, non-peer-reviewed, with the author's own
  training-data-contamination confound noted. A caution against reusing 2024-cohort numbers as current,
  not a replacement.
• Live-tracker caveat. Hallucinated-citation court trackers update continuously; a direct check on
  2026-08-13 returned 1,870 cases. Any hardcoded count needs a date.

Verification

• All 5 branches of truth_gate.sh's fallback ladder exercised individually — every one fail-open.
• stop_hook_active: true correctly allows (block-at-most-once respected).
• Block fires only when a completion claim and an unfinished marker co-occur; no false positive on
  an honest TODO sketch.
• Every statistic above traced to a primary source that was actually fetched.

Notes for review

• No Rust or skill content changed. Two files, both additive.
• One finding I did not fix, flagged for your call: hikmah validate returns "ok": true on the
  asymmetry above — it checks packaging, not behavior, so CI cannot currently catch a
  deterministic-gate-missing config. A behavioral consistency check between the two Stop adapters
  would close that.
• Separately: the Rust hook.rs keeps no internal block-once marker and depends on the host setting
  stop_hook_active. Worth knowing; not touched here.
• Opened for review, not for immediate merge.

Drafted with Claude Science.