adapter: gate the reversal exit cancel on candidate owners instead of every historical id - #563

… every historical id

A reversing entry cancelled the exits of every cohort id the run had ever
seen, which made each reversal cost grow with the number of distinct entry
ids. The cancel now visits only ids that can own a queued exit (pending
bracket legs, COOF requests, shadow rows, relative and anchored legs, live
exit rows), keeps the old sorted owner order, reserves the same scratch
handles through advance_handles, and keeps erasing empty bracket families.
Debug builds check every skipped id against the old per-store predicates.

A public-run test compares the gated path with the retained old loop on
five owner cases (live, empty, COOF, anchored, pending-bound): trades,
report and equity bytes, hash reads and reversal count match, and the
cancel-call count stays zero as distinct ids grow.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

The gate tested cohort membership with find(), which runs the implicit spill pass and restores a spilled row into the resident set. DiskResidentIdTable::contains() reads the index alone, so the gate no longer restores or spills any cohort row.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

test_private_fixture_reads pins a token at tests/CMakeLists.txt:3301. The two new registrations added two lines above it and failed that check. The reversal-gate row now replaces one comment line it folds, and the bulk-advance row joins the existing report-failure registration, so the file keeps its line count.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…n-bound manifest

The reversal exit-owner gate replaced the loop that built closed_cohorts, so its stale manifest entry and site records are removed. The gate's new local vectors (the selection scratch rows and the owner list) get open, unreviewed entries: the owner list is bounded by the queued exit rows, not by a constant, so nothing is marked bounded. The site inventory gains the matching records and the matches assignment.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ngine into reversal-exit-owner-gate

The reversal exit-owner gate adds two manifest entries (the scratch handle selection and the owner list) and removes one stale entry, so the unreviewed count is 1278 against a ceiling of 1277. The ceiling moves to 1278. No site, entry or bound changes.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…nt of the real-tree gate test

The real-tree gate test pins the number of open manifest entries. The manifest changed in this branch, so the pinned count moves to the number the gate now reports. No other change.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ling bound

Each applied reversal now checks its exit owners against the logical id table without restoring spilled rows. That probe descends a tree whose depth grows with the number of ids stored, so the 6000-to-24000-bar wall ratio rises a little above the linear 4.0: full release runs on the lab gave 5.02 to 5.06 against the old bound of 5.0, while the same run on main passed. The bound moves to 5.5, which still fails any super-linear growth (a quadratic run gives 16.0), and the comment says why.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… one place

The bound comment now gives the lab measurements (main 4.45 to 4.84, this branch 4.50 to 5.06), says that the log-depth id probe accounts for at most a factor of about 1.16 and that the rest is memory-hierarchy growth and noise, and that 5.5 still fails a quadratic run (16.0). The header line that still said "under 5" points at that comment. Comments only; no code or constant changes.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>