hooks: flag literal <DA> placeholder in the response closer

[zephyr91]

What

The 🗣️ closer's name slot must resolve to the DA identity file's name at render time. When substitution fails, the raw <DA> placeholder reaches the user undetected — FormatGate had no rule for it and DriftReminder's pre-write contract didn't measure it.

This adds the check to both hooks:

• FormatGate.hook.ts (Stop, observation-only) records a new da-placeholder violation to format-gate.jsonl.
• DriftReminder.hook.ts (UserPromptSubmit, pre-write) reports it in the next turn's FORMAT CONTRACT line, so the model is told before it writes.

Why closer-line only, and why last visible line

The check is anchored to the last visible line (the real closer), never to "last line containing the emoji". Two reasons it can't be looser:

1. Quoting the placeholder in prose — e.g. a PR or doc that discusses the format itself — is legitimate and a whole-message scan would false-positive on it.
2. A response that quotes the emoji in a body bullet (describing the format) must not be misread as a leaked closer. Anchoring to the final line, matching what FormatGate already does, keeps the two hooks consistent.

The check also requires the last line to actually contain the emoji before it treats a <DA> there as a leak, so a response with no closer at all isn't misattributed.

Verification

All four cases exercised against the new measure():

Input	Result
Real message that quoted the emoji + token in body prose	clean (no false positive)
Leaked closer (🗣️ <DA>: … as final line)	flagged
Correct closer (🗣️ <NAME>: …)	clean
Emoji + token in a body bullet, clean final closer	clean

No ~/.claude paths, names, or identity data in the diff; the {{PRINCIPAL_NAME}} scaffold is preserved.

[danielmiessler]

Ported the intent (both hooks had been restructured since): FormatGate now flags a literal <DA> in the closer's name slot only, credited to you, with a test. Next release carries it. Thanks, @zephyr91.