A test pins the run-cap park: it is noted and asked once per window, and a quoted reason is cut at a sentence #6927

test/crew-guard-cap.test.ts now pins the run cap's edges, and it found two real bugs in quoteCut, which I fixed in src/crew/guard.ts. tsc --noEmit, the new file and crew-guard.test.ts pass. I did not run the full check. prettier --check flags src/crew/guard.ts; I didn't try to fix that.

The bugs in quoteCut:

• Over the max: when no sentence end fell in the last 100 characters, the cut kept 600 and then appended …, giving 601. It now leaves room for the mark.
• Split pairs: the cut could end between the two halves of an emoji, leaving half a pair. It now backs off one unit when the cut would land inside a pair.

What the new test covers. It passes now in and sets timestamps by hand, so it never reads the wall clock.

• capReached:

- Under the cap is not capped, and at the cap is. - Runs older than the 24-hour window drop out. - Check-failure, conflict and review runs don't count, nor do retries, interrupted runs and skipped runs. - A director's word restarts the count; a reader's word doesn't.

• capNoted:

- The note is seen once written. - A reader's comment or a review after it starts no second note. - A director's word opens a new window with no note in it. - A note older than the window no longer counts.

• quoteCut:

- Short text comes back trimmed and otherwise unchanged. - Long text is cut at the last sentence end inside the final 100 characters. - Otherwise the cut ends in …. - The result never exceeds 600, and never splits a surrogate pair.

crew-guard.test.ts already covers the flow through queueAgentRun, so I left it alone. The new file calls the functions directly.

I did not touch recordCap.

Closes #6904

test/crew-guard-cap.test.ts now pins the run cap's edges, and it found two real bugs in quoteCut, which I fixed in src/crew/guard.ts. tsc --noEmit, the new file and crew-guard.test.ts pass. I did not run the full check. prettier --check flags src/crew/guard.ts; I didn't try to fix that.

The bugs in quoteCut:

• Over the max: when no sentence end fell in the last 100 characters, the cut kept 600 and then appended …, giving 601. It now leaves room for the mark.
• Split pairs: the cut could end between the two halves of an emoji, leaving half a pair. It now backs off one unit when the cut would land inside a pair.

What the new test covers. It passes now in and sets timestamps by hand, so it never reads the wall clock.

• capReached:

- Under the cap is not capped, and at the cap is. - Runs older than the 24-hour window drop out. - Check-failure, conflict and review runs don't count, nor do retries, interrupted runs and skipped runs. - A director's word restarts the count; a reader's word doesn't.

• capNoted:

- The note is seen once written. - A reader's comment or a review after it starts no second note. - A director's word opens a new window with no note in it. - A note older than the window no longer counts.

• quoteCut:

- Short text comes back trimmed and otherwise unchanged. - Long text is cut at the last sentence end inside the final 100 characters. - Otherwise the cut ends in …. - The result never exceeds 600, and never splits a surrogate pair.

crew-guard.test.ts already covers the flow through queueAgentRun, so I left it alone. The new file calls the functions directly.

I did not touch recordCap.