mirror of
https://tangled.org/evan.jarrett.net/at-container-registry
synced 2026-09-02 08:16:57 +00:00
Evan JarrettandClaude Opus 5
894dd243da
7d9de7c moved handle resolution below the sort-and-truncate, so a hold with
~500 crew stopped making ~500 serial identity lookups to render ten rows. It
shipped without a test, and the regression is a two-line move.
The property pinned here is the LOOKUP COUNT, not the wall clock. Timing would
pass or fail on how fast the machine is; the count fails precisely when
resolution moves back above the truncation. Mutation-verified: restoring the
old shape produces 60 lookups for 10 rendered rows against a 50-user hold, and
the failure message names the cause.
The second test covers the 3s resolve deadline the same commit added — a
stalled lookup must degrade to a bare DID rather than consume the reverse proxy
budget, which is what left the client hanging up mid-render before.
resolveHandle becomes a package var, since counting lookups is the only way to
observe either property from outside.
Two things worth knowing for the next test in this package. AdminUI.pds is a
concrete *pds.HoldPDS, so this needed a real one: NewHoldPDS with a file-backed
path (":memory:" is per-connection in libsql and disables the records index
QuotasByDID reads), then Bootstrap, or the first record write fails with
"cannot serialize undefined cid". And BatchCreateLayerRecords writes only to the
CAR store — the records index is fed from the repo event stream, which is not
running in a test, so BackfillRecordsIndex has to be called explicitly or the
quota query returns nothing and the count assertion passes vacuously at zero.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VwxF2N3HuZ8xSkx6nkirgB