fix(partition-keys): make claims nonblocking

[NikolayS]

Stack

This PR is intentionally stacked on #311 because both changes touch the partition lease implementation. Review only the commit above agent/fix-partition-lease-boundary.

What changed

• use a non-blocking row lock in claim_slot;
• return NULL when another transaction currently owns the slot row, allowing a first-free scan to continue;
• keep the distinct “slot is not subscribed” error when no row exists;
• reject non-finite lease TTLs;
• extend the real two-session lease harness with an uncommitted-row-lock contender;
• run that harness in every PostgreSQL matrix lane;
• add PostgreSQL 19 infinite-lease coverage and regenerate development artifacts.

Why

The API contract says a busy claim returns NULL; it previously blocked on SELECT ... FOR UPDATE. A slow claim/renew/receive transaction could therefore head-of-line-block a worker scanning for any free slot.

Fixes #309.

Validation

Red evidence on the parent/current behavior:

ERROR: canceling statement due to statement timeout
CONTEXT: while locking tuple ... in relation "partition_slot"

Green checks:

• real two-session harness: contender returns NULL behind an uncommitted lease row, existing lease-expiry/epoch recovery still passes;
• PostgreSQL 14: complete partition regression;
• PostgreSQL 18: complete partition regression + two-session harness;
• PostgreSQL 19 beta 1: complete partition regression including infinite-TTL rejection;
• bash build/transform.sh;
• workflow YAML parse, shell syntax, and git diff --check.

This is intentionally a draft and has not been merged.

[NikolayS]

REV rubric review

This is a manual execution of the repository's five applicable REV rubrics because the automated ultrareview quota was unavailable. It is not bot output. SOC2 was omitted.

• Bug/security: SKIP LOCKED plus the existence check distinguishes busy from absent slots without blocking; no additional defect found.
• Tests: the real two-session contender, expiry, epoch, and finite-TTL paths are adequate; CI is 17/17 green. Initial RED is retrospective.
• Docs/guidelines: final shell hardening adds the required safe IFS and reproducible psql form. The historical style: commit conflicts with the repo-specific type list.

Blocking patch findings: none. Merge-order constraint: this PR is stacked on #311 and must merge after it; rerun combined checks if the parent head changes. Exact real-user two-session evidence remains required. This COMMENT review is not an approval.

[NikolayS]

Real-user verification evidence for head 66dc74a6077367e994016a39fddcaa9b6eca591a, tested with its #311 parent.

I installed the stacked head into a fresh disposable database, then ran the real multi-backend harness:

PAGER=cat psql --no-psqlrc "$PGQUE_TEST_DSN" --set=ON_ERROR_STOP=1 \
  --file=devel/sql/pgque.sql
PGQUE_TEST_DSN="$PGQUE_TEST_DSN" tests/two_session_slot_claim.sh
bash -n tests/two_session_slot_claim.sh
shellcheck tests/two_session_slot_claim.sh
bash build/transform.sh
git diff --check

Against the parent behavior, the contender hit the harness timeout while waiting on an uncommitted slot-row lock:

FAIL: claim_slot blocked behind an uncommitted lease row
ERROR: canceling statement due to statement timeout
CONTEXT: while locking tuple ... in relation "partition_slot"

Against this head, the contender returned NULL promptly and the rest of the cross-backend lease lifecycle passed:

PASS: lease is exclusive across backends (outlives a dead one for its TTL); dead worker's slot recovered after TTL expiry; epoch bumped on takeover (fencing)

The harness uses distinct psql backends, including one holding uncommitted row DML, a contender, and a post-TTL successor. Focused partition tests passed on PostgreSQL 14, 18, and 19beta1, including the PG19beta1 infinite-TTL rejection. Shell syntax and ShellCheck also passed after the harness hardening.

RED/GREEN chronology: retrospective for the SQL behavior. The nonblocking implementation and regression were not published as separate test-first RED and GREEN commits. I independently replayed the harness on the parent to confirm RED and on this head to confirm GREEN. The later harness-only hardening did not alter the SQL behavior.

Full CI: 17/17 checks passed.

This comment records test evidence only; it is not a review or merge decision.