fix(inject-compartments): scope the injection cache to its source database

[iceteaSA]

Closes #307. Off clean master @ 885e93dc.

injectionCache is keyed on sessionId alone while holding database-derived content, so two independent stores sharing a session id replay each other's blocks. Cache entries now carry their Database handle; an entry whose db is not the current one is discarded instead of replayed, and every write records the db.

Reproduction (clean master, before the fix)

Store A seeds a memory and renders; store B is empty, same session id, defer pass:

PROBE A: injected=true  hasMarker=true
PROBE B: injected=false LEAKED_FROM_A=true

With the fix: LEAKED_FROM_A=false, and B correctly returns null.

What I deliberately did not touch

The defer-replay path is load-bearing for prompt-cache stability — it's what keeps <session-history> byte-identical across passes so a historian publish doesn't bust the prefix. The change adds an identity check before that path and leaves its behavior alone: same replay, same bytes, same conditions, for every entry whose db matches. Cache-stability tests (replays date-bearing m[0]/m[1] bytes unchanged on consecutive defer passes, the memories-only defer replay) pass unchanged.

I went looking for a cheaper fix first — clearing the cache in test teardown would paper over the symptom, but the invariant is a property of the cache, not of the tests, so it belongs at the cache.

Verification

plugin	3758 pass / 0 fail
pi-plugin	0 fail
typecheck	0 across 3 packages

RED-check: with the fix reverted, the new regression test fails with store A's marker present in store B's block — it genuinely guards the bug rather than passing vacuously.

bun run lint reports one pre-existing error in latch-permanence-guard.test.ts. That file is byte-identical to master here and fails identically on a clean master worktree under the lockfile-pinned Biome, so it's not from this change.

Honest scope

I found this because it broke test isolation downstream — two suites using ses-1 with separate temp databases, failing only when co-run. Session ids are unique per store in a normal run, so I'm not claiming a live-session failure and haven't constructed one. The claim is narrower: a cache keyed by session id shouldn't serve content rendered from a different database, and the check is cheap enough that the invariant seems worth holding regardless of whether a production path reaches it today.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Scopes the process-global injection cache to its source database and clears degraded re-anchor state on db mismatch. This prevents cross-database replays and early re-anchors when session IDs collide.

• Bug Fixes
  • Add db to injectionCache entries and call clearInjectionCache(sessionId) on a db mismatch to drop the entry and any degraded state.
  • Record db on all cache writes; same-database defer replay is unchanged.
  • Add regression tests to block replay across independent databases and to ensure no degraded-count inheritance for the same session ID.

^{Written for commit c6375ba. Summary will update on new commits.}

Greptile Summary

The PR scopes process-global compartment-injection cache entries to their source database while preserving replay behavior for matching database handles.

• Stores the originating database handle on every cache entry.
• Clears injection and degraded-reanchor state when a session-id collision crosses database boundaries.
• Adds regression coverage for injection replay and degraded-count isolation across databases.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
packages/plugin/src/hooks/magic-context/inject-compartments.ts	Associates cached injections with their source database and rejects cross-database entries before replay while retaining same-database behavior.
packages/plugin/src/hooks/magic-context/inject-compartments.test.ts	Adds regression coverage proving that an empty database cannot replay memory rendered from another database sharing its session id.
packages/plugin/src/hooks/magic-context/degraded-reanchor.test.ts	Verifies that degraded-pass bookkeeping is reset rather than inherited when the same session id is reused with another database.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Prepare compartment injection] --> B{Cached session entry?}
    B -- No --> E[Build from current database]
    B -- Yes --> C{Cached database matches current database?}
    C -- No --> D[Clear injection and degraded-reanchor state]
    D --> E
    C -- Yes --> F{Cache-busting pass?}
    F -- No --> G[Replay cached result]
    F -- Yes --> E
    E --> H[Cache result with current database handle]

Loading

_{Reviews (2): Last reviewed commit: "fix(inject-compartments): drop degraded ..." | Re-trigger Greptile}

[cubic-dev-ai]

All reported issues were addressed across 2 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[alfonso-magic-context]

Merged to master. Reviewed and verified independently: ran your regression tests on current master (71/71 on the two touched suites), re-ran the red-check myself (fix hunk reverted → the new cross-database test fails with store A's marker leaking; restored → green), and the full plugin suite (3,807/0). The clearInjectionCache-on-mismatch choice — dropping the degraded re-anchor bookkeeping with the block cache rather than leaving it to inherit across stores — is the part that makes this complete rather than cosmetic; the count gates a byte-changing re-anchor, so the leak's expensive form was the one you closed. Appreciated the honest scope note distinguishing the test-isolation repro from an unproven live-session claim; the invariant belongs at the cache regardless, exactly as argued. Rides the next release.