parity diff v0.5.1205 → main 2026-07-03: net +32/−10; 10 newly-failing tests need triage (incl. issue-945 guard confirmation)

[proggeramlug]

Summary

First full parity diff since the extended gate came back: v0.5.1205 tag run (2026-06-24, run 28110573641) vs current main edd608a44 (2026-07-03, run 28663806487, branch ci/parity-baseline-0703). Both parity-report artifacts attached to their runs.

Net: +32 newly passing, 10 newly failing, 63 failing in both (standing tail), 0 new-test failures. All 26 gap-suite failures in the current run are standing (old-numbered categorical gaps), i.e. the week's merge wave (#5466, #5854, #5866/#5867/#5871/#5885, the #58xx test262 wave) was a strong net positive on parity.

The 10 newly-failing tests (pass @1205 → fail @now)

test	status	note
test_issue_945_scalar_method_guards	parity_fail	independently confirms #5872's scalar-guard half as a real in-window regression (matches compile-smoke's ownMethodWrite scalarized guard failure; prime suspect #5466 by subject)
test_gc_aggressive_forced	parity_fail	GC behavior change in window — untriaged
test_edge_closures	parity_fail	untriaged
test_edge_interfaces	parity_fail	untriaged
test_generic_class	parity_fail	untriaged
test_issue_806_default_derived_constructor_forwarding	parity_fail	untriaged (derived-ctor forwarding — window includes capture/ctor-adjacent merges)
test_parity_argon2	parity_fail	was passing at 1205
test_gap_readline_3698plus	compile_fail	new compile failure
test_issue_3196_3198_tls_helpers	compile_fail	tls cluster — possibly runner env (also test_issue_3199_3200_tls_server_tlssocket, test_parity_tls, zlib ×2 failing in both runs or newly)
test_issue_3199_3200_tls_server_tlssocket	compile_fail	see above

Window: v0.5.1205 (c84bcb33b-era, 2026-06-24) → edd608a44 (2026-07-03). Each behavioral one needs a local run + bisect to attribute; they are small standalone parity fixtures.

Other findings from the same diff

• Parity runtime is NOT newly slow: the v0.5.1205 tag run's parity job also took 2h07m (current: 2h20m) — the "~20 min parity bill" note in test.yml is stale by months. The suite has 932 tests with 131 node_fails (node itself failing on the runner — likely missing native deps; these both dilute category percentages and plausibly burn serial timeout budget). Fixing the node-side failures and/or parallelism would give most of the time back.
• 13 skipped, 6 compile_fail total in the current run (tls ×3, zlib ×2, readline — clustered enough to smell like runner environment rather than perry; untriaged).

Suggested next steps

1. Triage the 6 behavioral newly-fails locally (run each fixture vs node on current main; bisect the window for any that reproduce).
2. Fold the scalar-guard confirmation into compile-smoke broken on main: issue-945 scalar-method IR guard fails (ownMethodWrite scalarized) — suspect #5466 #5872's investigation.
3. Separate pass on the 131 node_fails + tls/zlib/readline compile cluster (runner provisioning).

[proggeramlug]

Local triage of the six behavioral newly-fails (quiet machine, main @ v0.5.1227-era, fixtures vs node --experimental-strip-types, deterministic):

test	verdict	signature
test_gc_aggressive_forced	RECOVERED — now MATCHES node	plausibly fixed by #5937 (GC side-table scanners + static-field rooting, merged 2026-07-04)
test_parity_argon2	environmental	node itself fails locally (ERR_MODULE_NOT_FOUND: argon2 — fixture needs an npm install the runner had); not a perry regression signal
test_issue_806_default_derived_constructor_forwarding	real	super-args.log: seed=alpha:wrapped → perry prints undefined:wrapped — a user arg forwarded through a default derived ctor lands as undefined. Matches the 2026-07-02 audit §5 finding verbatim: no-own-ctor derived class inlining a capturing parent's ctor tail-splits USER args into cap slots when the child's capture union missed the parent's captures (lower_call/new.rs inline_constructor_param_values_with_class with caps_absent_from_args=false). Suspected root identified — needs the fix, not a bisect.
test_generic_class	real	TypeError: (string).getValue is not a function + exit 1 — a typed receiver mis-tagged as string; window and signature point at the #5466 representation-aware type-lowering wave
test_edge_closures	real	numeric output diverges (expected 1,2,3,4 block) — closure semantics; needs bisect in the v0.5.1205→now window
test_edge_interfaces	real	output diverges + perry exits 1 mid-run; needs bisect

Net: 3 real reproducers (one with the root cause already identified from the audit), 1 recovered, 1 environmental. The three real ones are deterministic small fixtures — good bisect candidates for the v0.5.1205→v0.5.1229 window.