fix(natives): heal better-sqlite3 in place so npm can't prune the heal away

[pacphi]

The symptom

ak sync contradicted itself in a single run:

✓ natives: …/ruflo/node_modules/agentdb: native installed
✗ still failing: [natives] 1/1 agentdb location(s) on WASM fallback (data-loss writes)

Both reports were true, ~30 seconds apart. The verifier was right; so was the installer.

Root cause

1. The ruflo@latest upgrade leaves the shared ruflo/node_modules/better-sqlite3 half-built — sqlite3.a compiled, no .node linked.
2. healNatives rung 1 installed better-sqlite3 into the agentdb location with --no-save. That produced a working local copy and verified true — honest at that instant.
3. Nothing in the tree declares that copy. When healAidefence ran npm install into the ruflo root, npm reconciled the tree and pruned it as extraneous.
4. Resolution fell back to the half-built shared copy → WASM fallback.

The same prune also removed agentic-flow/node_modules/agentdb — which is why the count read 1/1 where sync had just patched two locations. That 1/1 was the tell.

Worth noting: tests/kit/natives.test.mjs already had a test hypothesizing "an in-process npm install (aidefence) dedupes it away". That chased the resolver-cache symptom; this is the pruning root cause underneath it.

The fix

Heal the copy resolution already finds, in place, via the package's own prebuild-install || node-gyp rebuild — npm run is user-invoked and never gated by npm >=11.17 allow-scripts. Installing a copy is now the last resort, only when better-sqlite3 isn't resolvable at all. The heal lands on the declared copy, so a later reconciliation has nothing to prune.

Ordering alone would have been a half-measure: it fixes this sync, but leaves a landmine for any future npm install in that tree (ak x mcp, a later aqe install, or a user running npm by hand). Fixing the strategy makes the heal durable; the reorder is now defense in depth.

Also:

• Reorder natives after aidefence/aqe-solver in sync. Runs on security too — an aidefence install can wipe the binding even when the plan never flagged natives.
• Timeout 300s → 600s. A node-gyp sqlite3 compile is slow, and a half-built build/ is what a truncated one leaves behind. (Defensive — not proven to be a trigger here.)
• Inject runner into ensureNativeBsq3 so the ladder is testable without npm or a network.

Verification

Reproduced the original broken state (half-built shared copy + aidefence absent) → identical 1/1 … WASM fallback. Then with the fix:

• ak sync → ✓ natives: … native built in place → ✓ converged — no failing subsystems
• Binding lands on the shared declared copy; no extraneous copy planted
• Survives a subsequent tree-reconciling npm install — the exact operation that broke it before
• Second ak sync is a clean no-op (idempotent)

An earlier hypothesis (ordering alone) was falsified by testing — the binding survived. It only reproduced once natives ran first to create the extraneous copy. That's what pointed at pruning rather than ordering.

Tests

tests/kit/heal-natives.test.mjs — hermetic (synthetic fixture + injected runner; no npm, no network, runs on the full CI matrix). It simulates npm's prune-the-undeclared behavior and fails on the old strategy:

✖ heal survives an npm tree reconciliation (the sync convergence regression)
  AssertionError: STILL native after reconciliation — a heal that a later
  npm install prunes away is not a heal

3 of 5 fail on the old code, all 5 pass on the new. Gates: 87 mjs + 43 cjs tests pass, eslint exit 0, typecheck exit 0.