task room, landing and --one-model: real file names, one note sentence, every seat on one model, no build caches

[AbirAbbas]

Follows #1429 / #1485 / #1494 / #1436 review. One commit per item.

What was wrong, and what is true now

1. The work tab misread file names. It took a file's name from the last b/ in the diff --git line. So a person's x b/plandb.db was hidden as codeaf's own store, a name git quoted was drawn as octal escapes, and a leading space was trimmed. Now one reader (session.PatchSectionPath) reads the name the way git's own apply does, and both the filter and the room use it.
2. One sentence had two spellings. the worker reads a note at its next step was two constants. It is now one (session.RunNotePickupWord). A law test finds exactly one literal and checks that the manual quotes it.
3. --one-model did not reach a /task run's seats. The run engine seated the worker, planner, checker and probe from the profile's crew rows and CODEAF_CHECK_MODEL, which is how the tagged e2e billed claude-fable-5.1. Now every seat rides the conversation's live model.
4. A run's landing dropped files whose names git quotes. A name like odd name é'q.txt was dropped without a word. It read git status --porcelain without -z. Every reader of paths on the landing road now reads git's NUL form.
5. A landing committed untracked build caches (__pycache__/*.pyc in a repository with no .gitignore) and a worker's scratch copy. An untracked file matching one exact list now stays out: __pycache__/, *.pyc, .pytest_cache/, .mypy_cache/, .ruff_cache/, .DS_Store.
   • A tracked or worker-staged file still lands, and so do near-misses such as lockfiles.
   • The worker page tells the worker to delete its scratch files before plandb done.
   • The manual names the list.

Validation contract

• Every changed or added file is listed and landed under its real name (spaces, b/, quotes, accents, a newline). Only codeaf's own files are left out, judged on the real path.
• The note sentence is spelled once, and the manual quotes it.
• Under --one-model, the work, plan, check and probe seats all ride the live conversation model, whatever the crew rows and the environment say. Without the flag, nothing changes.
• An untracked cache on the list stays out of a landing and out of the work tab. A tracked or staged one lands, and so does every near-miss.

How it was checked

• Each item has a test that failed on dev and passes now. Examples:
  • TestLandKeepsEveryRealFilename (dev landed 2 of 6 names);
  • TestTheWorkTabDrawsRealGitFilenames;
  • TestTheNotePickupSentenceIsSpelledOnce (dev had 2 literals);
  • TestUnderOneModelEveryRunSeatIsTheConversationsModel (dev seated the profile's four rows and vendor/env-check);
  • TestLandLeavesOutUntrackedBuildCachesOnly.
• By hand, with the real binary on the engine road, running codeaf chat --one-model:
  • Every crew row was set to a different model slug. A /task wrote odd name é'q.txt, edited x b/plandb.db, and imported a module, which made a __pycache__ file.
  • Every call (106) rode deepseek/deepseek-v4-flash, including the review-round checker. Total spend was $0.013.
  • The landing commit held exactly odd name é'q.txt and x b/plandb.db. The .pyc stayed out.
• make pr-ready at box load 12–24.
  • The light gate is green.
  • In touched packages, the belt worker prompt's 16 KiB law failed on my first wording of the scratch-file sentence. That is fixed and rechecked after merging dev.
  • Every other red passes when run alone, or fails the same way on origin/dev:
    • internal/run: five supervisor/worker timing tests;
    • internal/session: TestPlandbCliWorkerOwnDone, TestAQueuedTaskThatSurvivedARestartStillHasItsBriefWritten and five others.

🤖 Generated with Claude Code