feat!: compact when the work moves on, not when a piece is finished

[wasd96040501]

What this changes

taskcut now compacts where the work moves on from a finished piece to another, not where a piece is finished. Hand over A, B and C in one message: it compacts between A and B and between B and C, and no longer after C. The end of a turn is no longer judged or compacted: the work is back with the person, and /compact before new work is theirs. Otherwise the structure is 0.8's: steps judged inside a turn, the turn ended, compacted, carried on with Continue..

One commit, on top of #6 (the $.model.complete reply-shape fix, merged): the question changes from "finished?" (DONE/WORKING) to "moves on?" (NEXT/SAME); the end of a turn is no longer judged; no step is judged when it says nothing, after a compaction until the model has changed something, or at a turn's first step (see below); the judge reads the whole conversation, as the permission classifier does. make eval-judge is added: labelled steps through the same call.

Why

• Compacting after the last piece helps nothing: nothing follows yet, and the person's next message is often about that piece ("also handle None", "commit that").
• A Claude Code 2.1.280 bug: a compaction after a request that ended on the person's words, /compact typed by hand included, answers them instead of summarising. taskcut ends a turn only after a step that followed tool results. 0.8 was exposed to it at every turn answered without a tool.
• The judge's prompt, per the classifier's source (2.1.280 bundle, xLt/Zke/iRr): the classifier reads its whole transcript and never leaves the oldest out; too long is no verdict. It is cached because it builds its own request with four cache_control breakpoints. $.model.complete (YTe) sends none, and the usage it reports is the API's own, so the judge is uncached today — no plugin-side fix without copying engine internals. The judge now reads the whole conversation too, so each prompt begins with everything the one before it read, ready for a call that caches.

How it was tested

steps	0.8 question	this PR
a piece done, another asked for	18/18	18/18
the last piece done	0/15	15/15
a piece in progress	12/15	15/15

(make eval-judge, 16 steps × 3, sonnet.) make eval-mechanism in a real interactive session, two four-task messages and a recall between: 9/9 checks, six compactions all between pieces, none after the last piece or at a turn's end, every file right.

Judge prompt, replaying 104 real steps: 15k characters on average, 30.5k at most; each prompt begins with the whole of the one before up to its step.

• make validate passes
• tsc passes against regenerated declarations
• Ran in a real interactive session (make eval-mechanism, eval driver on a pty)
• Claude Code version tested against: 2.1.280 (2.1.278's reply shape covered by unit tests only)

https://claude.ai/code/session_01Hy8i9XZbHT6uPuVZwAP8vT