An idle store is not read on every poll

[torstei]

Second of the changes from the full-read sweep: the polling loops. A follower polls every store it ships once a second, and most are idle on most polls. Each poll parsed the store's whole .rings (and serve() loaded its whole .grain) to find that there was nothing to send.

Change

• Shipper (ship.rs): a store is left out of the read when its cursor sits exactly at the end of the flushed chunks, no write-ahead entries are pending, and no .sap.seal handoff is in flight. The read also follows the live edge (query_entries opens a LiveTail for a store with a position), so a chunk-level test alone would be wrong for a wal store; both guards are tested. Anything unreadable is read, so the error surfaces. The stores returned by poll_excluding are the ones read, because take pairs the list with the buffer.
• Frames sender (ship_turn): skips serve() when the store has no chunk at or after the resume seq, from the index header and its last record (serve::nothing_from).
• serve(): no longer copies the selection twice (about 64 B per chunk of the store) before looking at max_chunks; selects by position instead. It reads only the grain pages it will send (grain::read_pages: length prefixes walked with a bounded buffer) instead of loading the whole grain, which used to happen even when nothing was selected.

Measured

Three stores of 300,000 chunks each, idle under timberfs feed --follow-from end:

	read per second	peak RSS
before	48 MB	128 MB
after	~0	7 MB

Tests

The skip: exactly the tape end, one byte either side, pending .sap entries, a .sap.seal, an unreadable store. A shipper that has caught up leaves both stores out of the read and picks the grown one up alone. Mutating the skip to always-idle fails 4 tests (it would lose data); to never-idle fails 2 (the optimisation would vanish). read_pages equals the pages load returns for every range, stops at a torn tail, and is empty past the end. A mid-store serve with max_chunks sends the pages of its own chunks.

Not here, and not verified

• query --follow's own poll still re-reads the index; it is entangled with the live-tail reconcile and the retention-overtaken note, so it is a separate change.
• A poll that has data still reads the whole .rings, and a serve turn with data still walks the grain from the front to its first page. Both want the reader-side index view planned separately.
• I could not confirm wake-up with a live feed run: the baseline binary does not deliver in my ad-hoc setup either (a consumer needs protocol progress reports), so that run says nothing about this change. The unit test above is the evidence.

Behaviour change

Shipper::poll no longer returns slices for idle stores. Nothing else consumes poll_raw / poll_excluding.

🤖 Generated with Claude Code