Engine plan tracker — GC correctness and performance (routing only; docs/engine-plan.md is authoritative)

[proggeramlug]

Hub for the engine effort — correctness and performance.

The plan is docs/engine-plan.md. This issue deliberately holds no content of its own: it links and routes, nothing more.

That constraint is the point. Every hand-maintained duplicate this project hit recently had silently rotted — gc_repsel_matrix.sh's header asserting copy-minor 12/22 when the true figure was 0/49 (#7255), a known_failures.json entry absent from the gap_snapshot.json the gate actually reads (#797), a README performance table with no generated markers. A tracker that restates the plan becomes the next one. If something here contradicts the doc, the doc is right and this is stale.

---

Part 1 — GC correctness

Shape: a GC-managed pointer exists somewhere the collector does not know about, across a point where the collector can run. Three homes, three mechanisms. Order is 0 → 2; layers 1 and 3 are independent and startable now.

Layer	Scope	Issues
0	in-process LLVM — the enabler	#7241
1	codegen lowering discipline	#7615 (campaign map), rfc-rooting-by-construction.md
2	emitted-code liveness (statepoints)	#7108, #7174
3	runtime Rust handles	#7231, #7254

Open instances (layer 2, subsumable — fix the instance, do not build new mechanism): #7246, #7247, #7248, #7251
Blocking bug: #7280 · Collector mode / RSS: #7056, #7161

Part 2 — Performance

Item	Issue
Array index range proof — measured 10.7× on matrix_multiply	#7286
method_calls is guard-bound, not call-bound	#7287
Build non-determinism (78 ms vs 3450 ms by file location)	#7288
Sound i64 specialization replacement	#7244
Write-barrier Phase B	#7187

Gates and instruments

#7282 (freshness fingerprint too broad) · #7272 (addr-class misses perry-ext-*) · #7255 (matrix arm liveness) · #7245 (native_proof_regressions unwatched) · #7261 (strip-dir leak) · #797 (skip-list provenance)

---

Two rules carried from recent experience

1. A green run that performed zero copying minors proves nothing (gc-matrix: four of the six PR-gating arms run ZERO copying minors corpus-wide (default/verify_evac/cons_scan_off at copy-minor 0/49; header still claims 12/22) #7255). Assert copied_objects > 0.
2. PERRY_GC_MOVING_LOOP_POLLS=1 at run time on a default-compiled binary is clean 10/10 and proves nothing (GC: #7154's residual is NOT fixed — the loop-polls config is red 0/30, and stock zod alone fails 5/40 #7280) — the poll codegen is not in the binary. Compile-time and run-time env must both be set. This is how an honest 30/30 gets reported on a broken build.