Test coverage follow-ups from the #3852 review (memory proof sizes, deleted and skipped tests, independent expectations) #3871

Part of #3868 (follow-ups to #854).

The independent review of #3852 found test-coverage gaps that do not change product behaviour. They were deferred here so #3852 could focus on correctness and the no-array send path.

1. Memory proof cannot see a latest-request leak. The retained-growth proof (journal-retained-growth*.test.ts) runs at a 30k-token context, about 100–150 KB. A leak that keeps one full copy of the latest request between turns stays under the 1 MiB allowance and passes, and so does a steady leak under about 1.7 KB per turn. Run the proof at two context sizes, with the larger one past 1 MiB, require the settled heap to stay flat across sizes, and add a trap that keeps only the latest request. The trap must fail.
2. Deleted regression test. compression-token-model-mismatch.test.ts (Refactor: Thread pending-content boundary explicitly through provider conversion pipeline instead of reverse-engineering it in ProviderContentEnforcer #2304) was removed in c008cd1. The logic it guarded is still there. Restore it against the source ladder.
3. Always-skipped trap. pending-caller-lifecycle.test.ts only runs its trap when CALLER_RETAINED_TRAP=1, which CI never sets. Make it run in CI (or turn it into an expected-failure trap like the R3 traps).
4. Self-referential expectations. About 80 hand-written boundaryRecovery expectations were replaced by a comparison against a production function, so nothing independent pins the expected results. Restore independent expectations.
5. Byte-check suites skipped on some PRs. The new byte-equivalence suites do not run on PRs that only touch core, agents or providers. Make the affected-shard selection include them.