Bounded resource acceptance matrix for #239

This is the continuation of existing PR #376, not a new PR. Previous failed receipts and checkpoint documents remain historical observations. Resource work stays on feat/239-provider-web-resource-campaign; parent review owns exact-head CI and merge. Qualified Angular T1 source 946f9b82 (including provider startup base bcd902cd) is merged as 990ed134 without changing those implementations. Three inherited browser documentation/descriptor conflicts retain the qualified parent versions. No SDK runtime Unavailable fix is made here.

Declared matrix before execution

The ticket specifies coverage, not 100000 deployments. These finite profiles were reported before execution. One owned container is limited to 3 CPUs, 6 GiB RAM with no extra swap, and 256 PIDs. Profiles run sequentially, not concurrently with another resource build. Cargo uses /workspace/target/phase3-239, inside the existing resource-only volume. The Angular target volume is mounted read-only for its maintained build and optimized compiler. No container, volume, failed receipt, previous binary or user directory is removed.

Profile	Cells / queue	Dormant deployments	Policy read owners	Churn	Deadline
smoke v2	1 / 2	4, 16 plus 3 fixture deployments	64	2 x 8 arrivals, 10 ms interval, 6 outstanding	240 s
campaign v2	2 / 2	4, 16, 32 plus 3 fixture deployments	96	8 x 24 arrivals, 5 ms interval, 8 outstanding	900 s
web-smoke v2	1 / 2	2, 4 actual Angular deployments	No capability policy owner	2 x 4 arrivals, 10 ms interval, 5 outstanding	600 s
web-campaign v2	2 / 2	2, 4, 8 actual Angular deployments	No capability policy owner	4 x 8 arrivals, 5 ms interval, 6 outstanding	900 s

Both web tiers use the existing protected external-capsule-v1 profile and 256 MiB per-cell memory ceiling, a 300-second isolated preparation ceiling, and five-second invocation budgets. Success, exception, explicit cancellation, client disconnect, overload, recovery, shared ingress and unrouting are distinct populations. An actual JS allocator heap counter is unavailable; OS RSS and configured Wasm memory ceilings must not be mislabeled as that counter.

Web v2 retains the failed v1 populations and deadlines but budgets two control jobs: one preparation owner and one read-only observer. Preparation retains at most 64 non-atomic samples, spaced at least one second apart, not a claim of exhaustive peak memory. Warm preparation verifies the actual in-memory prepared cache hit counter and unchanged resident entry/source/image/metadata totals; it is not mislabeled as an authenticated native disk-cache hit. The qualified SDK e07abf72 baseline is integrated as merge 70c41c09, with its root admission implementation unchanged. New runs identify that source separately. The parent's development reconciliation 5f38a9d5 is subsequently integrated as a6970f55; workspace version and lockfile changes require a fresh build.

The v1 smoke's requested 16-deployment population is retained in v2, not lowered to hide its failure. The unexecuted larger v1 profile's 64-deployment population is deliberately replaced by a new, smaller 32-deployment v2 profile: old and staged capability plans for 66 imported deployments would exceed the supported 128 ordinary policy read owners before observation headroom. No runtime maximum is raised. Configuring catalog entries alone did not budget the separate policy snapshot owner. The corrected runner preflights both generations and two observation owners (38 required for smoke, 70 for campaign), and records the actual configured ceilings. This is not proof that the old CLI exposed the limiting owner. A focused real policy-store regression tests refusal at the 33rd snapshot under limit 32 and admission/reclamation under limit 64.

Previous container exit

Before creating the new runner, docker inspect lsf-phase3-239-resource reported exit 137, OOMKilled: false, no daemon error, start 2026-09-19T16:58:05.456590369Z, finish 2026-09-19T18:15:30.615022441Z. Its command was sleep 28800, and container logs were empty. A bounded Docker event query returned no retained events. These observations do not identify who sent a kill signal and do not prove an application OOM or a resource leak. The old container remains untouched. The new lsf-phase3-239-acceptance container uses the same owned resource volume without overwriting previous outputs.

The first new build shell failed before Cargo launch because a login shell discarded the image's Cargo PATH. phase3-239-fixture-build-01.log is preserved; a non-login shell is used for the distinct build 02. That failure is setup, not product or acceptance evidence.

Evidence and acceptance boundary

New receipts use latent.phase3.resource-campaign.v2, exact binary/source, profile, configuration and signed-input identities, exclusive JSON/checksum creation, and the existing bounded/reaped process harness. Storage snapshots distinguish file count, logical bytes, unique inode bytes and allocated blocks; they include retained audit data and do not promise filesystem block dedup. Angular independently admits two packages/publications sharing one component.

PR #376 now explicitly targets development, not the former stacked SDK base. The first new executions below identify source 4b94e524; they are not evidence for a later merge of SDK #366 or for the eventual development merge commit. Any real Unavailable stops that profile and is retained for the parent's SDK/runtime diagnosis, without replaying mutations or claiming PASS.

Executed observations at 4b94e524

All three attempts used the same real node SHA-256 1724391b6cf80f21f2d8d6da5a0d317f5b67d602ada90bca47d7b70f83ecee6d and CLI SHA-256 308d28776ed6d47911fa1091b276c3d5b9c12805531f0fb145e28739665d7889. Each attempt retains its own build, fresh signed export, matrix and measurement receipt with original checksum sidecars. Times describe this shared Docker Desktop host only.

Profile / attempt	Actual observation	Disposition
Provider smoke v2 / 01	23 snapshots, all 4/16 added dormant deployments, two churn cycles, real HTTP/blob cold/warm/failure/cancel/recovery and clean shutdown in 32300 ms	Seven checkpoint checks passed; full ticket still pending
Provider campaign v2 / 02	67 snapshots, all 4/16/32 added dormant deployments and eight 24-arrival cycles; final HTTP recovery succeeded but blob recovery returned known guest-trap in 44305 ms	Failed resource-real-provider-output; node and peer force-reaped, not graceful shutdown
Web smoke v1 / 03	10 snapshots, all 2/4 dormant actual Angular deployments; cold preparation returned gRPC resource-exhausted after 117167543 ns, before any render	Failed resource-render-preparation-result; unknown RPC outcome, no preparation/render PASS

In provider smoke, dormant processes/threads/listeners plateaued at 1/8/1, with 34 open descriptors and RSS 62783488 bytes. After churn, active ownership returned to zero; six recovery samples retained 36 descriptors and RSS 81432576..81825792 bytes. These values distinguish warmed fixed resources from dormant density, not an allocator leak proof. Cold HTTP/blob elapsed times were 1157651469/1819662791 ns; retained calls include all outcomes rather than only successes. Three overload RPCs returned gRPC exhaustion with unknown outcomes, not proof of known queue rejection. Shutdown was clean but reported 11 blob stages, distinct from zero active blob handles/work.

The larger provider failure is consistent with the finite retained-stage contract: the maintained blob guest repeatedly writes the same four bytes, duplicate seal closes its writer but leaves an inactive stage for explicit privileged reclamation, and the standalone store defaults to 16 stages. No standalone reclamation operation was found in the configured provider path. The guest unwraps creation errors into traps. All blob arrivals in cycles 2..7 trapped, rather than demonstrating post-churn reusable write capacity. This source-based diagnosis is not a directly exported stage count from the failed node; that attempt has no graceful shutdown snapshot. Its last sample also reports 372 dropped audit observations, an explicit evidence limitation. Neither raising the production limit nor reducing the executed population is claimed as a fix.

The first web runner overlapped web prepare with an inventory RPC while workers.control allowed only one control job. Transport derives its control job ceiling from that setting and rejects excess jobs. This explains a harness self-contention risk consistent with the observed gRPC exhaustion; the generic RPC error does not itself identify the limiting owner. A corrected observation budget must be a new profile/receipt, never a relabeling of this failure.

No run above returned the parent's intermittent Unavailable. Web campaign, successful web preparation/render churn, standalone event/secret/child coverage, JS heap counters and OCI resolver/token/redirect campaigns remain gaps at this checkpoint. The earlier real HTTP/blob/secret/child Rust fixture observations are separate from standalone acceptance.

Protected compiler observation failure

Web smoke v2 / 04 at 324f7110 retained 15 samples, including the real compiler child, before PermissionError aborted collection after 6175667650 ns of preparation. The compiler deliberately enables Linux NotDumpable; no ptrace capability, security-profile relaxation or alternate unprotected compiler is introduced. The corrected sampler records unreadable protected-child descriptors/network metrics as unavailable, including unavailable aggregate counters and explicit sample denominators. It still retains readable process/thread/RSS observations, rejects unavailable root-node descriptor inspection, and never substitutes zero. An exiting child or unreadable descendant list marks tree totals incomplete. This failure remains immutable and does not establish preparation success.

Historical results: checkpoint, attempt ledger, methodology.

Running the matrix

tools/phase3_resource_campaign.py --matrix prints the complete finite matrix without starting a process. tools/run_phase3_resource_acceptance.py builds and selects the exact Cargo fixture executables, verifies their nonempty listings, records the current native build, and exports fresh real-clock signed inputs immediately before each profile. A failed profile stops the matrix. All reports, including preparation failures, use new paths; no mutation is replayed.

export CARGO_TARGET_DIR=/workspace/target/phase3-239
export CARGO_BUILD_JOBS=3 CARGO_PROFILE_DEV_DEBUG=0 CARGO_PROFILE_TEST_DEBUG=0
python3 tools/run_phase3_resource_acceptance.py --revision FULL_SOURCE_COMMIT --output /workspace/target/resource-attempt-UNIQUE --guest-capsules /workspace/target/phase3-resource-guests-02 --angular-build /angular-target/phase3-angular-t1-build/actual-selected-01 --compiler /angular-target/phase3-226-release/release/latent-aot-compiler --host-condition shared-docker-desktop-host

--profile smoke selects only the first small provider tier; each invocation still needs a fresh output directory. Ordinary CI runs the small Python and automatically discovered Rust regressions. The real matrix requires the explicit workflow_dispatch input run_phase3_resources, default false, independently of the unchanged run_catalog_scale heavyweight gate. Manual CI creates its own guest inputs, uses the job's actual Angular build and optimized compiler, and uploads even failed JSON/checksum attempts. No hosted CI execution is claimed by this integration change.