feat: add neural-gate — behavioral constraint echo detection

[YuhaoLin2005]

What

A post-session check that verifies whether behavioral constraints actually appeared in agent output. Complements filesystem gates (which verify file state) with NL-channel verification (which detects whether rules shaped behavior).

Why

Filesystem gates check that files were updated, hooks are wired, and exit codes fired. But they don't answer: did the rules we wrote actually change what the agent did? Neural-gate fills this gap by scanning session transcripts for constraint keyphrases.

How It Works

Stop hook → read constraint keyphrases from config → scan session transcript → report echo rate per constraint → never blocks (detection only, not enforcement)

Honest Limitations

• Keyphrase echo ≠ behavioral compliance (mentioning a rule ≠ following it)
• False positives: agent can echo a constraint while ignoring it
• False negatives: constraint may shape behavior without being explicitly named
• Explicitly a weak proxy; v2 roadmap: semantic similarity scoring

These limitations are structural — the same Prose Barrier (verification and generation share the NL channel) that makes filesystem gates necessary.

Integration

Module D of a composable quality pipeline. Works standalone. Complements:

• pre-task-calibration (feat: add pre-task-calibration — mechanical pre-action verification #1422): pre-action verification
• delivery-verification (feat: add delivery-verification — session-end quality gate #1423): session-end file state verification
• persona-review (feat: add persona-review — multi-persona adversarial review skill #1420): adversarial depth

The only gate that explicitly does not block — it detects, reports, and lets the human decide.