Heap-snapshot leak-regression harness

[truthixify]

Tier: M (2-4 days) | Type: test

Context. test/leaks/scan-leak.test.ts (PR #114) already writes a raw heap-final.heapsnapshot via v8.getHeapSnapshot() (lines 62-72) for CI debugging and asserts a regression slope on process.memoryUsage().heapUsed (line 113). What it does NOT do is diff retained-object counts by constructor across before-and-after snapshots, which is the failure mode we most fear after PR #154 (pipelined scan) and PR #156 (batch-path optimiser) landed: a retained WeakRef / buffer pool that leaks slowly. This issue adds the constructor-level diffing layer, publishes the artifact from the nightly job, and documents the harness.

Scope.

• Extend test/leaks/ with a runner that takes structured v8 heap snapshots before / after N scans and parses retained-object counts by constructor.
• Diff constructor counts, fail on growth over a configured threshold.
• Bench on Node with --expose-gc so GC is deterministic.
• Nightly job (not per-PR) publishes the diff as an uploaded artifact.

Acceptance.

• Harness produces reproducible retained-object counts across two runs on the same commit
• Regression injected in a spike branch fails the harness
• Nightly CI publishes the diff as an artifact
• Doc'd in test/leaks/README.md (new)

Files. test/leaks/heap-snapshot.test.ts (new), test/leaks/README.md (new), .github/workflows/ci.yml.