feat(sparq-lws-core): sparq-lws-core: wire per-resource WAC into notification subscribe/receive (the

[sparq-orchestrator]

🤖 SPARQ agent

What / why

Automated implementation of the trusted task in #6071, routed to sparq-rust-impl on
Claude Opus 5 (alias opus5, provider model claude-opus-5).

Fixes #6071

Local gate

• Policy profile: crate-scoped
• Result: passed before push

Merge posture

DRAFT — pending cross-provider review. Publish never arms; arming happens ONLY in the registry
review-fix approve path (arm_auto_merge=true), gated on an opposite-provider approve
verdict with ci-summary / gate as the objective backstop.

Performance (local, release build, 4 cores, oha -z 10s -c 32 over loopback)

This branch changes subscription authorization only, not the read or write paths. These are the Solid routes, three alternating runs per binary, with SOLID_SERVER_SEED_BENCH=100:

Solid route	main 2e90b5a (rps / p99)	this PR 124a18f (rps / p99)
GET /bench/public/doc	64,963 · 64,867 · 63,948 / 1.54-1.61 ms	64,718 · 63,003 · 65,506 / 1.50-1.59 ms
GET /bench/listing/ (100 children)	36,407 · 35,811 · 36,871 / 3.01-3.03 ms	36,142 · 37,223 · 38,628 / 2.82-3.00 ms
PUT anonymous (401 path)	33,739 · 32,008 · 33,880 / 2.20-2.41 ms	34,479 · 34,257 · 35,225 / 2.02-2.39 ms

No difference beyond run-to-run noise.

---

Generated by Claude Code

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 1 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 2 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 3 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — the autonomous review loop parked this PR: the review round budget is exhausted at 3 round(s) (base 3, hard cap 6) with no extension left — the top fix tier has run, the latest verdict does not grade the PR improving, and no pushed fix at or above the pinned floor awaits re-review; a human must decide

This is the MACHINE-owned capacity park (review:parked), not a human question: it remains a DRAFT and will not be auto-armed. A human can re-admit it by removing the live machine park label(s) — review:parked here and status:parked on the source issue (whichever are present; a needs:user unlabel on either surface also opens the budget window) — the budget restarts from the latest gesture.

[sparq-orchestrator]

🤖 SPARQ agent — automatically re-admitted this MACHINE capacity park: its own starvation cause can no longer be observed, and the fleet is demonstrably healthy.

This park is older than the rolling model-health window, so whether the specific condition that parked it has cleared is NOT provable any more — leaving it would make an automatic hold a permanent one. Instead the fleet has recorded sustained successful runs across multiple accounts with no launch failure, the most recent at 2026-09-04T07:16:58Z (evidence fleet-health/fleet/5f5b2dc3fb8e231b/33846793471.1 — provider/account-fingerprint/run from the model-health window; no raw handle). That is a HEURISTIC about fleet health, not proof that this park's own cause cleared. The machine park label(s) are being removed and the review loop re-admitted with a real budget window.

This consumes that evidence EXACTLY ONCE: the same evidence can never re-admit this PR again, a later park needs FRESH evidence that has not been consumed (for the cause-recovery route, a new outage-and-recovery pair), and at most 2 automatic re-admissions are ever granted to one PR — past that the loop stops and asks a human. A human hold (review:needs-user / needs:user) is never touched by this path, and neither is a park a human applied by stamping one of those human-owned labels. A park a human applied by stamping the MACHINE-owned soft hold is re-admitted here ONLY when this bot's OWN park-reason receipt already classified the episode class=capacity — the machine never clears a park it never classified. To place a hold no machine may lift, use review:needs-user / needs:user.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 4 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 5 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 6 recorded.

[sparq-orchestrator]

🤖 SPARQ agent — the autonomous review loop parked this PR: the review round budget is exhausted at 3 round(s) (base 3, hard cap 6) with no extension left — the top fix tier has run, the latest verdict does not grade the PR improving, and no pushed fix at or above the pinned floor awaits re-review; a human must decide

This is the MACHINE-owned capacity park (review:parked), not a human question: it remains a DRAFT and will not be auto-armed. A human can re-admit it by removing the live machine park label(s) — review:parked here and status:parked on the source issue (whichever are present; a needs:user unlabel on either surface also opens the budget window) — the budget restarts from the latest gesture.

[sparq-orchestrator]

🤖 SPARQ agent — automatically re-admitted this MACHINE capacity park: its own starvation cause can no longer be observed, and the fleet is demonstrably healthy.

This park is older than the rolling model-health window, so whether the specific condition that parked it has cleared is NOT provable any more — leaving it would make an automatic hold a permanent one. Instead the fleet has recorded sustained successful runs across multiple accounts with no launch failure, the most recent at 2026-09-04T22:50:46Z (evidence fleet-health/openai/dc2d751954080e13/33926991000.1 — provider/account-fingerprint/run from the model-health window; no raw handle). That is a HEURISTIC about fleet health, not proof that this park's own cause cleared. The machine park label(s) are being removed and the review loop re-admitted with a real budget window.

This consumes that evidence EXACTLY ONCE: the same evidence can never re-admit this PR again, a later park needs FRESH evidence that has not been consumed (for the cause-recovery route, a new outage-and-recovery pair), and at most 2 automatic re-admissions are ever granted to one PR — past that the loop stops and asks a human. A human hold (review:needs-user / needs:user) is never touched by this path, and neither is a park a human applied by stamping one of those human-owned labels. A park a human applied by stamping the MACHINE-owned soft hold is re-admitted here ONLY when this bot's OWN park-reason receipt already classified the episode class=capacity — the machine never clears a park it never classified. To place a hold no machine may lift, use review:needs-user / needs:user.

[sparq-orchestrator]

🤖 SPARQ agent — cross-provider review round 7 recorded.

[jeswr]

ci-fast is red only because its three jobs (lint, test, conformance) were cancelled before running. GitHub Actions has an open incident where hosted runners are not being assigned. No step of this PR failed. The same steps are now running locally against current main, and the PR will be re-run once runners recover.

---

Generated by Claude Code

[jeswr]

Local ci-fast run while GitHub Actions runners are down (local-gate rule). Head 48e5e79 merged with main bfe8514:

• clippy (core crates, -D warnings): pass
• nextest (core crates, --profile ci): pass, 3016 passed, 27 skipped
• doctests: pass
• W3C SPARQL conformance: pass, 1225 pass + 4 documented divergences (ratchet 1229)

48e5e79 also fixes three findings from an independent review of this PR: the notification gate now applies the opt-in ODRL deny exactly as GET does; open sockets re-check access before every frame and close with 1008 on revocation; and the receive-side test now drives the receive route itself. Benchmarks are in the PR body.

---

Generated by Claude Code

[jeswr]

Local gate addendum. The gate passed on head 48e5e79 merged with main a496ec9 (clippy, nextest, doctests and SPARQL conformance 1229, all green).

Skipped range a496ec9..d2c96e0: one file, research/noir-bounded-vec-capacity-assert-5027.md. It is docs only, no ci-fast inputs, and is not pulled in by any include_str! under crates/. Merging on that basis under the coordinator's updated rule 4.

---

Generated by Claude Code