unimem raw-access dependents — baseline diagnostics

2026-10-10. Only committed parent honeycrisp 7b148906263da5ea8aabc5ddb0e2e44ff2288ecc is used; no candidate overlay or consumer source migration. This is local dependency evidence, not a release, compatibility approval, warning-free result or full ring closure.

Reproduction inputs and commands

scenarios.json pins every extracted repository, revision, archive SHA256 and committed lock SHA256. Archives are the 19 exact-origin inputs catalogued in /tmp/unimem-access-origins.knYGlE/archives.json; their acquisition and registry relationships are in /tmp/unimem-block-access-dependent-plan.md. prepare.nu, prepare-local.nu, and prepare-drift.nu create separate sibling roots from those archives. No working-tree source is an input. The latter two scripts were explicitly authorized for local lock diagnostics after recording the origin-locked failures; they do not edit source/manifests or the origin roots.

commands.json and logs/<scenario>/<label>.json record each actual Cargo command, cwd, explicit environment, times, exit, and lock hash before/after. Adjacent .stdout/.stderr retain complete output, including metadata's full source paths and every warning. run.nu returns successfully when it has written a receipt; the Cargo exit is the receipt's exit_code, not the wrapper's. Initial wrapper parser errors were corrected before Cargo began; they are not compiler/test outcomes. Commands execute rustup 1.95.0 Cargo/RUSTC/RUSTDOC, except Nika's declared nightly-2025-11-26. tool-environment.json pins versions. Private targets use two build jobs and disabled incremental compilation.

Every scenario first attempted metadata --format-version 1 --locked --offline. Missing cache artifacts were fetched only with fetch --locked, followed by locked offline metadata. metadata-paths.json verifies every local package stays inside its scenario and records the actual driver manifests. Nox's local closure has no honeycrisp package; its successful rung does not exercise unimem.

Observed results

C below means cargo check --workspace --all-targets --locked --offline. All named tests/checks likewise preserve their scenario lock and run offline. Warnings remain visible and were not promoted to errors or suppressed.

Lock treatment and immutable-source proof

Origin scenarios retain every committed lock byte-for-byte. Bita/glia origin scenarios still have no lock. Local locks are retained under retained-locks/, with hashes and complete diffs; reuse these exact files for future candidate comparisons, without a second resolution.

• Bita/glia generated one new lock each with generate-lockfile --offline.
• Mir/erga started from the committed lock and ran metadata without --locked: the only diff is removal of the stale libc dependency from unimem's entry.
• Nox likewise retained its seeded graph: its own package version changes 0.1.2→0.2.0, plus seven existing path packages update to their pinned source versions (Hemera, lens-core and five strata packages). No registry versions change in this diff.

These five scenarios are explicitly local-resolved, never origin-locked or release candidates. source-comparison.json contains full diff -qr receipts against the exact extracted originals: the only differences are those five root Cargo.lock files. failing-source.json also pins the unchanged files behind observed code failures, including Glia's failing integration fixture.

Test bounds and remaining gates

Bita filtering is intentional: seven *throughput* tests plus qhash_lc_bench::tests::attribution are benchmark workloads. The GPU throughput test alone traverses batches through 2^26 nonces, and attribution repeats 200,000 hashes per phase after warmup. The bounded run keeps the independent 16-nonce GPU expectation comparison and 256-nonce easy-hit test. Its eight excluded names:

metal::mining::tests::gpu_mining_throughput
miner::tests::concurrent_throughput
sha256_mb::tests::sha256_throughput_compare
qhash_lc::tests::lc_throughput
qhash_lc::tests::lc_throughput_parallel
qhash_lc_cpu::tests::cpu_throughput
qhash::tests::qhash_throughput
qhash_lc_bench::tests::attribution

Perla's excluded tests::bench_gpu_tile exercises 768×512×2176 panels with batches 1/4/8 and repeated timing passes. Its retained jackpot oracle uses 16×16 panels and k128. Both Bita and Perla test code can return early if Metal setup fails, so the specific oracle was repeated with --exact --nocapture in test-gpu-observed. Both repeated single-test runs pass with no skip output; the source guards print on setup failure. The silent-return easy-hit test alone still does not establish hardware execution. No global driver safety claim follows.

Zoya cache tests instantiate 16,776,896-byte caches (actual constant, despite stale 256MB comments) and 1,024-item test datasets, not the mining DAG. Xena's pair/quad fixtures use bounded 543,744-byte scratchpads. Erga tests use 4,096-row synthetic tables and local unsigned wallet fixtures. Its declared release build used RUSTC_BOOTSTRAP=1 cargo build --release --locked --offline; the isolated cargo run -p erga-cli --bin erga --release --locked --offline -- difftest then passed on Apple M4 Max, comparing 512 nonces over a 4,096-row table. The explicit CLI argument selects only the offline oracle. No live wallet transaction, network pool, miner, packaging, installation or large epoch-table build ran.

Remaining unexecuted commands are not counted green: full Bita/Glia/Xena/Mir workspace tests; Perla GPU benchmark; Zoya debug suite; unavailable Nika/Trisha/ Cyb closures and product/platform gates; strict Clippy/format/semver gates for these downstream repositories. Candidate validation is not authorized here. Re-run the same locked scenarios only after the owner's candidate source freeze.

At final observation private build targets occupy 5.2GiB, below the 30GB limit; no cleanup was necessary. SHA256SUMS covers the compact receipts, helpers, retained locks and full log streams; build products and source archives are not part of that receipt payload. No repository file, commit, PR or candidate changed.