unimem raw-access dependents — candidate diagnostics
2026-10-10. This is the local diagnostic overlay of the frozen Block-access
implementation on honeycrisp parent 7b148906263da5ea8aabc5ddb0e2e44ff2288ecc.
Source manifest SHA256:
1ead683ed4e75cfb0c2edd196e6e18e979f3ca83634e69c3bcdd2bfe72e1b1be.
The native source worktree is /Users/master/cyber/honeycrisp-kadek-block-access.
No consumer source was migrated. Compatibility and ring closure are red.
Reproduction and integrity
scenarios.json records every origin revision, archive and committed lock.
prepare.nu re-extracted the same 17 baseline scenarios and reused the same
five retained local locks, without a new dependency resolution. All source
matched the parent baseline before the explicit 16-path Honeycrisp overlay;
15 scenarios close over Honeycrisp. The two Nox scenarios have no Honeycrisp
package, so their unchanged parent check/test evidence is reused only as
owning-rung evidence, never candidate driver coverage.
Baseline /tmp/unimem-dependent-baseline.cX82Mt/SHA256SUMS is pinned by
d2f7485eb0e59b06025b3301e53e5d23bc05ea35fa005656325b6718f16db5d2.
Candidate overlay.json pins every source byte. nu execute.nu ran the 37
selected commands sequentially with identical baseline argv, locked/offline,
private targets, two jobs, incremental disabled and exact Rust1.95.0 binaries
(Nika uses the same declared nightly2025-11-26 as parent). Tool executable
hashes and version comparisons are in toolchains.json.
logs/<scenario>/<label>.json records command, environment, time, actual Cargo
exit, source-overlay SHA and unchanged lock hashes. Adjacent stdout/stderr
are complete and SHA-pinned, including warnings. Wrapper success means only
that evidence was written. nu verify.nu independently checked all receipts,
the baseline payload manifest, lock equality, all successful metadata local
paths, final 17 source trees and exact changed-path whitelist. comparison.json
retains emitted error/warning headers, source locations and test result lines
for each command on each side; it does not treat matching exit codes as proof
of matching diagnostics. final-source-comparison.json retains full tree diffs.
The first verifier attempt expected hashes in parent command receipts; those
actually live in the parent payload manifest. It stopped before writing any
verification artifact and was corrected to verify that manifest directly.
verification.json: 37 actual commands, 16 exit0 and 21 nonzero. These include
metadata and repeated build configurations, not 37 independent test suites.
No source/manifests/locks outside the explicit Honeycrisp overlay changed.
Causal comparison
All commands below map to exact argument arrays and revisions in the receipts.
C means the selected all-target check, with Metal/GPU features where named.
| Scenario | Parent | Candidate and interpretation |
|---|---|---|
| Xena | Whole-workspace C fails E0608 in xena-bench/src/gpudebug.rs:192,195; scoped hash/reference C and 4 tests pass. |
All selected C/tests fail compilation: two E0133 sites xena-hash/src/scratchpad.rs:34,42 call newly unsafe as_bytes_mut. The previous E0608 paths and later GPU calls are masked, not fixed. No candidate tests run. |
| Bita local lock | Metal C, bounded hash5 tests (62 passed, 8 benchmarks filtered), explicit GPU oracle pass. | All selected C/tests fail at the changed API: metal/mining.rs emits 13 E0133 and 5 E0596 findings in C. Four shared-receiver methods and a non-mutable local cannot request exclusive mutation; raw imports/views now require safety proofs. Candidate GPU oracle never executes. |
| Erga local lock | C and bootstrap release/difftest pass; workspace tests reach 11 passes and two existing chime overflow failures. | C/tests/release/difftest now fail compiling blake-bench/src/lib.rs: E0133 at99/107/114, E0596 at107. Earlier runtime failures are masked. No candidate runtime/difftest result exists. |
| Mir local lock | Default/Bevy C fail missing Gpu::sync at src/gpu.rs:32. |
Same emitted E0599; no tests and no claim that masked paths are compatible. |
| Glia local lock | Workspace C fails E0063 missing has_attn_bias at run/tests/tier5_graph_ir.rs:22; library C/tests pass. |
Same workspace error; library C/tests pass again, 92 tests. Existing warnings remain. |
| Zoya | C and workspace release tests pass. | Same checks pass, 27 tests. No debug-suite claim. |
| Perla | GPU C, bounded workspace tests and explicit no-capture GPU oracle pass. | Same checks pass: 7 bounded workspace tests, bench_gpu_tile filtered; repeated exact GPU oracle passes once with no setup-skip output. That repeated oracle is not an additional unique test. |
| Mir/Erga origin locks | Locked metadata requires existing lock update. | Same prerequisite failure; local-resolved diagnostics above remain separate from origin. |
| Bita/Glia origin | No committed lock; locked metadata fails. | Same prerequisite failure. No origin-locked build claim. |
| Evy | BBG feature resolution fails. | Same missing backend-unimem requirement. |
| Nika/Trisha | Required exact vendor closure unavailable. | Same missing chaff/tasm-lib manifest; no moving or dirty vendor substitute. |
| Cyb | Authoritative Nu origin unavailable; isolated closure deliberately incomplete. | Same first missing staged cybergraph manifest. This is a setup limitation, not proof cybergraph is absent upstream. |
The API breaks are intentional removal of unsafe operations exposed as safe,
not baseline failures. They require consumer migrations with initialization,
owner lifetime, CPU/GPU exclusion and completion/visibility proofs; inserting
unsafe mechanically would not discharge those obligations. No consumer fixes,
version/pin edits, releases or compatibility approval are part of this unit.
Public semver evidence is supplied separately by the native implementation.
Warnings remain in full logs. comparison.json preserves their emitted sets;
compilation stops earlier in Xena/Bita/Erga and can mask old warnings or errors.
No global zero-warning or unchanged-diagnostic claim follows. Unavailable
Cyb/Nika/Trisha/Evy closures, full downstream suites, cross-platform checks,
format/Clippy/semver of consumers and Kelvin drift are not counted green.
README.pre-execution.md and preparation.json intentionally preserve the
historical pre-overlay state, superseded by these completed receipts.