fix(driver-diagnostic): exclusive owner-only publication, and delete six unreachable items

[schickling-assistant]

The defect

driver_diagnostic::atomic_json was the only one of the ten publication helpers in the tree that used File::create on a predictable temp name in an agent-writable directory — while its sibling in delivery_ledger documents, thirty lines of prose away, exactly why each of those matters. Four defects in five lines:

	before	after
staging open	File::create — no O_EXCL, follows a symlink and truncates its target	create_new — refuses with AlreadyExists
mode	0o666 & ~umask, so world-readable	0o600
staging name	.driver-diagnostic.<pid>.tmp — one name per process, two writes to one agent dir interleave	.driver-diagnostic.tmp-<pid>-<counter>
directory entry	never synced	best-effort sync_all, exactly as delivery_ledger does

Pinned by driver_diagnostic::tests::a_planted_staging_symlink_is_refused_and_the_record_is_owner_only, which plants a symlink at a staging path, asserts the refusal and that the victim's bytes survive, then asserts the published record's mode and that no staging residue is left.

The parent-directory sync stays best-effort here deliberately: this is a defect fix and must not also change the helper's success/failure contract. Making the sync strict across every helper is T1-FSAT's job.

T1-DEAD

Six deletions, each backed by an exhaustive search over src, tests and crates:

• pretrust::pretrust_codex — 0 callers; its own doc says reconciliation does not call it. pretrust_codex_at is the live entry.
• context::append_decision — a one-line forwarder with no production caller (the CLI already calls append_decision_to_dir). Its 5 unit-test callers now call that directly.
• agent_author::resource::add_resource — an 8-argument forwarder superseded by add_resource_with_selector, which production already uses. Its 17 unit-test callers pass the None the forwarder passed, so the tests exercise the same call.
• identity::AddressBookEntry::bus_address — 0 callers. Every other bus_address in the tree is AgentSpec::bus_address(this_host), a different method that takes an argument.
• ding::observed_poke — the repo's only #[allow(dead_code)] in src, superseded by observed_poke_with_window.
• resource_profile::relative_path — 0 callers anywhere. reports/deepclean-st2.md filed it as merely test-only; it is dead.

Two tests that could not fail

• hooks::tests::omp_launch_classification_is_exact carried a duplicated #[test], which made the harness report it twice — visible in the gate's test count, which is why --lib reads 743 here and 743 on main while this PR adds a test: +1 real test, −1 phantom.
• Above it sat the doc comment belonging to pi_launch_classification_is_exact — the test whose missing #[test] refactor: pure moves — test tails, the clap surface, and eight submodules #501 restored. The two defects are one mistake: a test was pasted in and its attribute and doc landed on its neighbour. Attribute de-duplicated, doc moved onto the test it describes.

Visibility narrowing: only where it is warning-free

CodexHoldReason and CodexTerminalError → pub(crate); codex_app_server::{load_current_binding, load_current_control_state} → #[cfg(test)] fn.

reports/deepclean-st2.md also lists CodexRuntime, CodexThreadBinding, CodexControlState, CODEX_CONTEXT_VERIFIED_VERSION, CodexObservedState, resource_profile::last_commit and flapping::is_parked. Each was tried and reverted, for a reason worth recording:

• The first four wake dead_code on their test-only accessors the moment crate-external reachability stops suppressing the lint. That is why they are pub today, and an #[allow(dead_code)] to hide it would be a worse trade than the wide visibility.
• CodexObservedState appears in a pub method of the still-pub CodexControlState, so narrowing it trips private_interfaces.
• flapping::is_parked has two integration-test callers (tests/run.rs), which link the lib as an external crate — narrowing is a hard E0624, not a warning. The report's row is wrong.

Checks

• Gate nix build .#st2 --no-link -L: EXIT=0 (pty st2.clean-4.gate, log /tmp/st2-clean4-gate.log).
• Test-name set diffed against refactor: pure moves — test tails, the clap surface, and eight submodules #501's gate run: exactly one addition (the new symlink test) and one removal (the phantom duplicate). No other test moved, renamed or vanished.
• cargo check --all-targets: 0 errors, and no warning that is not already on main (resource_profile_supervisor, ding/mod.rs, tests/resync.rs).

Deferred, with the analysis attached

T1-FSAT, T1-LOCK and T1-DUP are not in this PR. Each was scouted to an applicable apply-sheet — module source, per-caller mapping, per-site behavioural delta, test plan — and each carries at least one deliberate behaviour change that deserves its own reviewable PR rather than riding inside a "dedupe" diff:

• T1-FSAT tightens six callers' record mode from 0644 & ~umask to 0600 and makes delivery_ledger's directory sync strict; it also uncovered a second, distinct defect (harness_state::persist_floor stages at the fixed literal .harness-state.seq.tmp, so two writers for one agent can rename a torn floor into place).
• T1-LOCK's premise in the report is wrong: CatalogLock is a domain lock, not a file-lock primitive, and its own module doc says state-plane traffic deliberately does not use it. The correct shape is extracting the transport into src/flock.rs, which also hardens five lock files with O_NOFOLLOW.
• T1-DUP changes the wire-visible (None, None) exit label from "exited" to "exit unknown" (unreachable for a reaped child on Linux, and pinned by nothing today) and moves omp's failure diagnostic off the agent's inherited stderr.

Tracked with the full apply sheets in the follow-up issue.

Posted on behalf of @schickling

field	value
agent_identity	dev3.direct.omp.43sz6ujq
session	dev3.43sz6ujq
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.1.7
agent_runtime	OMP 18.1.7
tooling_profile	dotfiles@39a19af