[Known limitation][v1.0.0] O_TMPFILE successful-open identity divergence leaves observation incomplete (fail-closed) #195

Status and purpose

KNOWN V1.0.0 OBSERVER / RAW-V2 CAPABILITY LIMITATION — DOCUMENTATION ONLY; NO REMEDIATION AUTHORIZED BY THIS ISSUE.

Record the verified existing limitation for future ExecSurface engineering without changing any existing release, implementation, baseline, completeness rule, source evidence, or qualification/merge gate. This is a future engineering consideration, not a declaration that v1.0.0 has failed an independent end-to-end workload test, a newly proven security vulnerability, or that fail-closed behavior should be removed.

Verification of applicability to v1.0.0 (2026-11-09)

• Public immutable v1.0.0 release tag resolves to source commit e70169b959f2163715090c371335fa6c5591e3e4.
• In v0.1.0-alpha.6, v1.0.0, and current main, the source file crates/execsurface-observe/src/linux_ptrace.rs resolves to the same Git blob: 6ec1076673b93577e2125e5e0e8abafa9445c2fb.
• The shared source includes side_effectful_open_identity_divergence around line 1288 and recognizes O_TMPFILE among side-effectful open flags around line 1407.
• On a successful side-effectful open, a lexical path can differ from the kernel-resolved returned FD path (e.g. lexical /tmp versus /tmp/#<inode> (deleted)). Raw observation v2 lacks a dedicated successful-open object identity for that case. The observer emits a divergence warning and marks the observation incomplete instead of authorizing an unsafe canonical surface.
• This is a verified source-level applicability to v1.0.0. The specific external workload failure described below was executed on Alpha.6, not v1.0.0. No v1.0.0 end-to-end reproduction of the same original workload is claimed.

Preserved external evidence — never conflate runs

Original, 2026-10-08, Alpha.6: external Actions run 37745730944, evaluator's report.

• hippoley/Residual-Necessity external workload; GitHub-hosted Ubuntu x86_64; Python 3.11; python -m pytest -q.
• Target workload 15/15 PASS; ExecSurface doctor PASS; learn exit 2, raw observation incomplete; no trusted baseline, no downstream check or drift verdict.
• Workflow run HEAD: ca7d14f937ce9069dc2eb742d20b7a03751ee073; checked-out PR merge ref: 1283968a7d1797c45a47de5ee3c3aaa2cc399dbf (distinct source identities).
• Original run exposes no downloadable Actions artifact at the time of review; the original run's syscall flags and exact causation are unproven.

Separate later diagnostics — not a rewrite of the original:

• Diagnostic report and run 37764233781: incomplete_capability, side_effectful_open_identity_divergence, raw_observation.complete=false, 42 typed effects; reviewed artifact supplied explicit O_TMPFILE flags for two opens. This run also had a separate missing jsonschema dependency issue.
• Later corrected diagnostic run 37869646669: independently reported 41 effects.
• Another run 37922133708 had 142 typed effects in its separately inspected artifact, and the diagnostic workflow ended SUCCESS while the observation remained incomplete and the downstream check was skipped. Workflow SUCCESS is not an ExecSurface PASS.
• Source/artefact audit on Draft PR #190 identifies bounded EXPECTED_FAIL_CLOSED / KNOWN_RAW_V2_CAPABILITY_BOUNDARY for this diagnosed mechanism, not for all possible incompleteness.
• External intake and scientific-review authority #140, P8 #114, and Draft PR #190 retain the source evidence and qualification gates. These remain separate from this v1 technical limitation tracking record.

Guidance for a future explicitly authorized engineering cycle (not a request to execute now)

1. Re-verify current tag/branch/blob and scope. Reproduce with the actual observer on a supported Linux filesystem, using both an O_TMPFILE scenario and an ordinary O_CREAT negative control; keep the original Alpha.6 external report immutable.
2. Study a semantics-preserving way to represent verified successful-open kernel-object identity and alias information, including whether raw-v2 can remain stable or requires a separately approved versioned representation. Do not silently reinterpret existing evidence or change source authority.
3. Preserve conservative fail-closed: if object identity/completeness remains ambiguous, no baseline approval or PASS/REVIEW/BLOCK may be manufactured. An unsupported O_TMPFILE filesystem is an unsupported/failed test, not a passing test.
4. Require targeted Linux regression/negative-control and independent review under the project's active governance before any implementation, contract migration, release statement, or merge.

Boundaries / non-claims

• No claim that v1.0.0 was externally reproduced as failing on the original workload, that all O_TMPFILE uses fail, that a fix already exists, or that an unsafe outcome was issued.
• No adoption, endorsement, privacy-redaction completeness, product defect exoneration or general readiness claim.
• No code change, no new branch, no PR modification, no release action, no external communication, and no execution authorization in this issue. Future product engineering should discover this issue and independently verify the current state before proposing a specific fix.

Decision: RETAIN AS DOCUMENTED KNOWN CAPABILITY LIMITATION; REMEDIATION DEFERRED UNTIL SEPARATELY AUTHORIZED.