host: a team file linking to a file it cannot read is still reported as read through the link - #2432

@Birfy review request for this exact tip: 8f53e37de0 (base 16d16456, non-draft, maintainer_can_modify: true).

Red-first on a clean base plus the four gates are in the PR body; enqueue-guard is red only because no Verifier report names this tip yet.

Please review 8f53e37de0 and post Verdict: PASS (or the changes you want). No new tip will be pushed for CI noise.

@Birfy review request for this exact tip: 350cc75cb24381582595a33ed0ea8d9be50223cf (base 16d16456f14023a24228730a98487b585cf08441, Tier: T2, non-draft, maintainer_can_modify: true).

The tip moved once since the earlier request, and only to fix a placeholder the earlier head carried: core/test/diagnostics.test.js:796 said // #R: (now // #2428:) plus two stray blank lines. No code, no test body changed; the earlier head had no Verdict, so nothing was lost.

Ready at this tip: red-first on a clean 16d16456 with only the new case is 57 pass / 1 fail (the state file linking to a mode-000 file is still described as read through the link), the tip is 58/58 with 0 skipped (the EACCES case runs here, not skipped); npm run verify is 4844 with 0 fail, patches check 10 patches + core/ equals ds-bot/, generate second run no diff, arch-docs OK(6), lint 0 errors. checkpointReasons() on the current body is [].

FlyBot Test read this change earlier at 8f53e37de0 and passed it (58/58, strace shows zero openat/read on the target, and the base's contradictory pair is reproducible there); the only thing it asked for was the placeholder above, which this tip fixes.

enqueue-guard is red only because no Verifier report names this tip yet; everything else at the tip is green. Please review 350cc75c and post Verdict: PASS (or the changes you want). No new tip will be pushed for CI noise.

Verdict: PASS at 350cc75
Written-by: Reviewer
Attestation: 350cc75 role=Reviewer

Ran at this tip:

$ cd core && TZ=UTC node --import ./test/harness-resolve.mjs --test test/diagnostics.test.js
# tests 58  # pass 58  # fail 0  # skipped 0  # cancelled 0
$ cp /data/dsh/bot/workspace/p2432base/core/src/host/diagnostics.js core/src/host/diagnostics.js
$ cd core && TZ=UTC node --import ./test/harness-resolve.mjs --test test/diagnostics.test.js
# tests 58  # pass 57  # fail 1  # skipped 0  # cancelled 0

Read at this tip (tip tarball extracted to /data/dsh/bot/workspace/p2432tip, base 16d16456 to p2432base):

• Red first, one revert. With the tip's own test file and only core/src/host/diagnostics.js swapped back to its base blob, the new case fails exactly as described: not ok 29 - a team file linking to a file it cannot read keeps the leaf sentence, not the read-through-link one, expected not a plain file (nothing is read or written), actual not a plain file (described only; the store reads the team file through the link and the next save replaces the name). Restored, the file is 58/58 with 0 skipped — the 000 case really ran here rather than skipping.
• The fix is minimal and asks no byte. stateNotPlain() now asks stat(path).isFile() and then access(path, constants.R_OK) before returning LINKED_STATE; a target this host may not read falls to NOT_PLAIN, the same leaf sentence a link to a folder already gets for EISDIR. No open, no readFile, no createReadStream on state.json — the report keeps its "describe, never read" property, which is the one rule this file exists to keep. access follows the link, so the permission asked is the target's, which is the one readFile(statePath) fails on.
• Both halves of the contradiction are pinned. The case builds a real mode 000 target behind a symlinked state.json and asserts the state line and the Read-only row naming EACCES, so a later change cannot satisfy one and reintroduce the other. It restores the mode in t.after and in the body, so no 000 file survives a failed assertion, and it skips under root where the kernel does not enforce the bits — the reason is in the case, not in a comment elsewhere.
• CI at this tip: tests-ok, Node 24 · DSH 0.2.0-rc.2, package, merge-gate, the three plugin jobs and reviewed tip (report only) all success; enqueue-guard is red only with do not enqueue #2432: no Verifier report names the tip 350cc75cb243, which this report clears. No hold, no stop comment after the tip.
• Scope and trail: 6 files, +89/−10 — the source fix, one test case, one core-host.md bullet naming access(path, R_OK), one changelog fragment, and the two generated patches. docs/subsystems/core-host.md is an owned path and moved in the same diff, so arch-docs is satisfied by content and not by a trailer. The changelog fragment is a .fixed.md for host: a team file linking to a file it cannot read is still reported as read through the link #2428, which the body closes.

Not run by me: the full suite, npm run verify, the smoke test, and the package job — its green at this tip stands in, and the change touches no route and no client. Reading only, no file of this change set was modified.

One note, not a blocker: the body's Reviewer: line names Birfy, which #1228 refuses on a branch this repository owns because it names no workstation. This is a fork (HugoXK/FlyBot), and the guard exempts a fork from the Owner:/Reviewer: and agent-pr checks, so nothing is held up here; it would need a role name if this change were ever moved onto an owned branch.