test(router): pin the tag-store registry LRU order after a touch - #1247

devcorner: rango-t158 · session rango-s285

Problem

The explicit tagged-store registry (registerExplicitTaggedStore in packages/rangojs-router/src/cache/cache-policy.ts) is LRU-bounded at 64 entries: re-resolving a store does set.delete(store); set.add(store) so a module-singleton store moves to the most-recent end and is not evicted by per-request churn. The test named for that touch, "keeps a re-resolved (live) store reachable despite per-request churn (LRU touch)", still passed with set.delete(store) removed. Without the touch, churn evicts the live store, and the next resolveCacheStore(liveStore) adds it back, so the final has(liveStore) holds either way. The touch had no test that pinned it (found in review rango-t131-c2; predates S1).

Test-only; no router code changes.

Before / after (internal shape)

Before: 200 rounds of "resolve the live store, resolve a fresh store", then expect(set.has(liveStore)).toBe(true).

After: fill the registry to its cap with s0..s63, re-resolve s0, resolve one newcomer, and assert the whole order by store identity:

expect([...explicitTaggedStores].map((s) => names.get(s))).toEqual([
  ...stores.slice(2).map((s) => names.get(s)), // s2 .. s63
  "s0",
  "newcomer",
]);

The stores are mapped to names because distinct empty MemorySegmentCacheStore instances are deep-equal, so toEqual on the instances themselves would not see the order.

Code changes

• packages/rangojs-router/src/cache/__tests__/tag-store-registry.test.ts: the churn test is replaced by "moves a re-resolved store to the most-recent end, so eviction takes the least recently resolved (LRU touch)". The other six tests are unchanged.

Test plan

• vitest run src/cache/__tests__/tag-store-registry.test.ts: 7/7 pass.
• Red under the mutation: with set.delete(store) removed from registerExplicitTaggedStore, the new test fails (1 failed, 6 passed): expected [ 's1', 's2', … ] to deeply equal [ 's2', 's3', … ], i.e. s0 was evicted and s1 kept. Restored, it passes.
• The base version of the test file under the same mutation: 7/7 pass, confirming the gap.
• Router tsc --noEmit (its include covers src/**, tests included): pass.
• node tools/check-internals.mjs: OK, counts unchanged (request-context-object-create 21, internal-vi-mock 327, rsc-param-literal 71, browser-import-into-server 1).
• oxlint --deny-warnings: 0 warnings, 0 errors. oxfmt --check .: clean.
• test:unit:all and e2e not run: test-only change in one file, run directly.

Practices

• Lint: oxlint --deny-warnings and oxfmt --check . on the whole repo; nothing found.
• Dead code: the change replaces one test body and adds no helpers or exports; nothing is left unused.
• Tests: the touched file passes; the new test is shown red against the mutation it targets and green without it.
• Code review: the diff was reviewed by hand. The registry cap (64) is spelled in the test with a comment naming EXPLICIT_STORE_REGISTRY_CAP, as the neighbouring bound test already does with the literal; exporting the constant from cache-policy.ts for a test was left out to keep this change test-only.
• Docs: none describe this test; nothing needs changing.

Notes

• The removed churn scenario's claim (a live store stays reachable) also holds through re-adding after eviction, so it did not pin the touch. The new test pins the mechanism itself; the bound test above it still covers growth under churn.