[santoshkumarradha]

Idle Factory polls no longer read every job document on each look. Held: do not merge until the owner has reviewed before-and-after numbers and authorized it. Stacked on #2358; retarget after that lands.

What changed

With #2358 applied, the largest remaining engine work on a real install with 355 jobs is store.List. A read-only sample of the idle engine shows 135 samples in List: lease Pull 82, job settlement 35, runner capacity 12, OverLimit 11. Each look decoded every document.

• List now stats every file on every look and decodes only files whose identity, size or time changed. It returns a deep copy, so callers, migration and output retention never touch the kept document.
• Lease grants and execution gates still read the job under its lock.
• List reads every document again at least once a minute, so a decision never rests on file times for longer than that.

Measurements

Read counts on a floor of 300 jobs, unchanged floor inside the 60 s bound, from the tests:

Route	Before	After, unchanged	After, one job changed
Pull	900	0	1
OverLimit	900	0	1
Settle and capacity	1800	0	1

Real app, Mac Studio (M3 Ultra), 355 jobs, 3 job chats open, about 4 minutes after start. Two runs each, same desktop build, same session, including the 60 s re-read. Not the laptop. Before is 56d8b5acb (#2358 without its re-read); after is 4aa44a98f (this PR and #2358's re-read). The later merges of app/v3 were not re-measured.

	Before	After
Engine CPU used at about 3:26	16.0, 15.5 s	14.1, 14.4 s
Idle, three 20 s windows	8.9, 8.9, 3.7% and 8.5, 8.9, 3.6%	7.3, 8.0, 3.5% and 7.5, 8.7, 3.5%
Memory (RSS)	240, 241 MB	254, 258 MB
Key to paint, p50 / p99	24.8 / 40.2, 24.6 / 38.6 ms	24.7 / 40.6, 24.8 / 39.7 ms

The gain on this machine is small: about 10% less engine CPU for about 14-17 MB more memory. Typing does not change. The target is the laptop, where this work is a larger share.

Laptop before and after: not measured yet. It needs an owner-approved install. This PR stays draft until those numbers are here.

Correctness tests

Case	Test
Another process's Stop, Start, budget change, create, delete, take, release, rename	TestPullAndCapacitySeeExternalFloorChanges, TestSettleSeesExternalStopStartAndBudget
Fresh authority after selection	TestPullRechecksLockedJobAfterListing, TestSettleRechecksStopAndBudgetAfterListing; the lease fix itself landed separately in #2362
Ownership, retention, migration	TestListOwnsMutableDocuments, TestListReappliesOutputRetention, TestListReappliesRepositoryMigration
Replacement with the same size and time	TestListSeesRenameWithSameSizeAndTime
Looks racing writes, under -race	TestFloorLooksRaceExternalWrites

Limits

• The kept documents add memory, since they hold the output that retention trims from what callers see. Measured in the pending rows.
• Every look still stats every file. Only decoding is avoided.
• Every look still trusts file identity between full reads. An in-place rewrite that keeps the same inode, size and modification time can wait up to 60 s. List reads every document again at least once a minute, the bound, tested by TestListSeesInPlaceSameTimeEditWithinItsBound. Grants still read under the lock.

🤖 Generated with Claude Code

[AbirAbbas]

Closing because this is already on dev. Every behaviour this PR adds landed through other PRs after app/v3 reached dev (#2561, #2597):

• Unchanged files are not re-read. Factory job reads go through one process-wide cache that decodes a document only when its file identity, size or time changed, and returns a deep copy (cachedRead and cloneDoc in internal/factory/store/readcache.go, from factory: a run-delivered PR can merge; an idle engine stops re-reading its floor (r5 E10, E3) #2512).
• The lease desk reuses unchanged queue items. The desk's Pull and OverLimit reuse unchanged queue items and their order (floorQueue in internal/factory/lease/queue.go, from engine: held leases stop re-reading the floor; one group commit for closing chats; clock leaves with its window (r7) #2517 and Idle factory workers wait for changes instead of re-decoding the store #1931).
• Idle settlement skips an unchanged floor. Job settlement skips the look entirely when the floor is unchanged (floorUnchanged in internal/factory/serve/jobs.go, from Idle factory workers wait for changes instead of re-decoding the store #1931 and engine: an idle floor settles from memory, held jobs stop re-reading, a clean quit hands leases back (r8) #2519).
• Its base is on dev too. The open job-chat stop watch this PR was stacked on is on dev through An idle job chat stops reading every job ten times a second #2358.

The branch stays as it is, so nothing is lost.