Cleanup/hardening: fix five open correctness/security defects, adopt ADR-340 invariants

[ruvnet]

Summary

One hardening pass over the open defect backlog: five issues fixed with regression tests, the underlying invariants recorded as ADR-340, and three new issues filed for defects found during the work (#930, #931, #932).

Closes #825, closes #901, closes #903, closes #907, closes #908.

Fixes

#825 — ruvector-delta-index: DeltaHnsw insert hang (root-caused: self-deadlock)

The hang was not an unterminated traversal: connect_node's prune branch held a write lock on a neighbor node while prune_neighbors re-locked the same node through self.distance(). parking_lot locks are not reentrant, so the inserting thread deadlocked — but only when a neighbor list overflowed m0, which the unseeded test RNG made nondeterministic. Reproduced deterministically (seeded, 400 inserts, dim 8 — hung pre-fix, <1s post-fix).

• Neighbor's vector now comes from the held guard via a split borrow; prune_neighbors drops self-loops before measuring distances.
• greedy_search carries an explicit progress bound (strictly-decreasing distance ⇒ ≤ nodes.len() moves) so a malformed graph terminates loudly.
• Tests seeded; prune-heavy regression test added; ruvector-delta-index rejoins the core-platform CI shard (checklist item from ruvector-delta-index test_insert_and_search hangs indefinitely (DeltaHnsw insert/search) #825).

#901 — ruvector-tiny-dancer-core: NaN confidence panicking route()

Non-finite score/uncertainty now rejected at a single choke point covering both the gated and ungated paths, tripping the circuit breaker (matching the VoI path's contract); the decision sort uses total_cmp as defense-in-depth. Regression test poisons the model weights with NaN (the issue's corrupted-model scenario) and asserts error-not-panic.

#907 — ruvector-graph: strong-Arc pool (unlinked-database reuse + unbounded growth)

Ported ruvector-core's post-#902 reference shape wholesale: Weak per-path slots, Drop-under-the-slot-guard eviction, slot reaping. Regression tests: erased-path recreation observes an empty database; reopen-after-last-drop preserves data and reaps the pool entry; 40-iteration barrier-synchronized concurrent drop-vs-open probe, zero tolerated failures. The non-canonicalized-key tail note from #907 is split out as #932 so it survives this close.

#908 — ruvector-context: require_private_root owner check

The root must now be owned by the process euid in addition to being 0700, via rustix::process::geteuid() (no unsafe, already in Cargo.lock through fs4 — one line of lockfile churn, no new transitive dependency). As the issue argued, this rejects only roots that already failed later with a bare EACCES.

#903 — harness: locale-dependent benchmark cache key

canonical() in benchmark.ts now sorts keys by code unit (RFC 8785), with the Intl.Collator-stub regression test copied from the #890 fix shape. The same defect in research.ts (feeding sha256/embeddingSpaceId digests) is fixed in the same way.

⚠ One-time cache invalidation (called out per the issue): changing the ordering changes the benchmark cache key, so existing .metaharness/cache entries no longer match and benchmarks recompute once. This is the accepted cost of a locale-independent key and the reason this change ships in its own commit path rather than riding a feature PR.

ADR-340 — Correctness-Hardening Invariants for Hot-Path Primitives

Records the six invariants these fixes instantiate so the defect classes stop recurring: total float orderings on untrusted values, no lock re-entry (read through the held guard), the Weak+Drop pool shape as the only sanctioned pool pattern, identity-not-just-mode filesystem trust checks, code-unit canonical encodings for anything hashed, and seeded RNGs in tests of nondeterministically-triggered behavior. docs/adr/INDEX.md updated minimally (generator's full regeneration churns every row's git date on a fresh clone; node scripts/adr-index.mjs --check passes).

Follow-up issues filed from this pass

• ruvector-delta-index: delete() of the entry point leaves search anchored to a phantom node — the rest of the graph becomes unreachable #930 — DeltaHnsw::delete() of the entry point leaves search anchored to a phantom node; verified empirically: the whole graph becomes unreachable and search returns only id="".
• Sweep: 91 panicky float sorts (sort_by(partial_cmp().unwrap())) across 27 crates — NaN-reachable in core query paths (ADR-340 invariant 1) #931 — inventory of 91 remaining sort_by(partial_cmp().unwrap()) sites across 27 crates, triaged user-reachable-first (ruvector-core query paths, ruvector-postgres backend-abort risk).
• Storage pools key on non-canonicalized PathBuf — two spellings of one file produce two Database::create calls and a redb lock error #932 — both storage pools key on non-canonicalized PathBuf (two spellings of one file ⇒ redb lock error), split out of ruvector-graph: connection pool still holds strong Arcs with no eviction (unlinked-database reuse + unbounded growth) #907.

Validation

• cargo test -p ruvector-tiny-dancer-core — 49 passed
• cargo test -p ruvector-context — 44 passed
• cargo test -p ruvector-delta-index — 15 passed (pre-fix: hung; suite now <1s)
• cargo test -p ruvector-graph — full suite green
• cargo clippy over the four touched crates — clean (pre-existing workspace warnings only)
• harness: npm test — 167 pass / 0 fail (the one prior failure was a missing local Python jsonschema, unrelated)
• node scripts/adr-index.mjs --check — OK, next available ADR-341

🤖 Generated with claude-flow

https://claude.ai/code/session_01L5ffi8NiKAQNabK3D5t2gK

---

Generated by Claude Code

[ruvnet]

CI note: Build Tiny Dancer linux-arm64-musl is red here, and the failure is not this PR's.

It is the known toolchain drift from #900: the link of third-party redb fails with unsupported linker arg: --fix-cortex-a53-843419 before any first-party code is reached. This PR surfaces it only because it touches crates/ruvector-tiny-dancer-core/, which re-triggers the path-filtered workflow that hadn't run since its last green pass on 2026-08-03.

Root cause is now pinned down empirically (details in the fix PR): stable 1.98.0 (2026-08-18) began emitting the Cortex-A53 erratum flag for aarch64-unknown-linux-musl, and no zig release accepts it — verified against zig 0.13.0, 0.14.1, 0.15.2, and 0.16.0. The failure reproduces with a minimal cdylib link on 1.98.0 + zig 0.13.0 and disappears on 1.97.0 with the same zig.

The fix is #934 (pin toolchain: "1.97.0" beside the existing zig 0.13.0 pin), kept as the separate CI PR that #900 explicitly asked for rather than folded into this diff. Once #934 merges, this check goes green here on re-run; every other failure signal on this PR (the initial Rustfmt red) is already fixed on the current head.

---

Generated by Claude Code

[ruvnet]

CI note: Tests (core-and-rest) was cancelled at the 240-minute cap, and this failure is not this PR's.

It is exactly the pre-existing #928 compile-size problem: the job ran 02:53→06:53 UTC and was killed at the cap during compilation, the same outcome #928 documents on PR #926's run 32702332738 (cancelled mid-compile of ruvector-temporal-tensor-wasm, 16/17 other jobs green) — filed the day before this PR existed. This PR's diff adds only test code and small fixes to crates in that shard, nothing that moves a four-hour compile wall.

Not re-running: the cap-out already reproduced on an independent PR (#926, per #928), so a re-run would spend another four runner-hours to confirm what is already confirmed. Not fixing here: #928 explicitly reserves the shard split/caching change for maintainer review as a CI-policy decision, so porting a workflow change into this PR would preempt that.

Signal coverage note: everything this PR actually changes is tested by green shards — core-platform (which now includes ruvector-delta-index again, passing in 7 minutes), core-and-rest-heavy, core-and-rest-wasm, both Clippy jobs, Rustfmt, and the harness job all pass on this head. The two remaining reds are both pre-existing and tracked: this one (#928) and Tiny Dancer arm64-musl (#900, fix in #934).

---

Generated by Claude Code