Fix variant_surfaces double-clone in reconcile_with_typed_cache

[gunbai-bot]

Summary

Isolated Fix B for whole-tree compile speed (bold-crane-271): in reconcile_with_typed_cache, stop holding two live Rcs across rc_map_insert by building variant_surface first, then moving variant_surfaces into the insert.

Previously each module slot did rc_map_insert(variant_surfaces.clone(), …, build_variant_export_surface(…, variant_surfaces.clone(), …)), keeping refcount ≥ 2 and forcing Rc::make_mut structural copies on every im::HashMap update in the reconcile fold.

M1 cross-batch resolve memo is already on main — not in this PR.

Benchmark (local, compile_clean_diagnostic_histogram, whole-tree)

Variant	HISTOGRAM_ELAPSED_SECS	reconcile gantt
main (302e955b78)	188.5	150.0s
B-only (this PR)	192.8	152.6s

Run-to-run variance on this box (~2%) swamps the micro-fix; the change removes a proven refcount≥2 call-site pattern regardless.

Test plan

• cargo fmt --all --check
• compile_clean_diagnostic_histogram whole-tree — HISTOGRAM_STATUS ok, HISTOGRAM_TOTAL_HARD 0
• CI build job green
• CI floor (ci) — pending at mark-ready time

[gunbai-bot]

Verified review 40762 against f96c323 — finding holds, no code change needed.

What changed: reconcile_with_typed_cache now builds variant_surface first, then moves variant_surfaces into rc_map_insert instead of rc_map_insert(variant_surfaces.clone(), …, build_variant_export_surface(…, variant_surfaces.clone(), …)).

Semantics: build_variant_export_surface only reads the prior map (no in-place mutation); computing the surface before insert sees the same pre-insert state as before. Net effect is one fewer Rc<im::HashMap> clone per module slot in the reconcile fold — the insert path no longer holds two live Rcs across Rc::make_mut.

Verdict: Agree with APPROVE — narrow mechanical cost-shape fix, identical behavior.

— sent from bold-crane-271

[gunbai-bot]

Verified review 40769 against f96c323 — finding holds, no code change needed.

The diff matches the described shape exactly:

• build_variant_export_surface reads the pre-insert map via variant_surfaces.clone() at cli_run.rs:6645
• variant_surfaces is then moved into rc_map_insert at 6648–6649 (no second clone on the insert path)

build_variant_export_surface is read-only over surfaces (v1_compiler_infer.rs:16070+); reordering compute-before-insert preserves behavior while dropping one Rc clone per reconcile slot. Agree this is a valid §6 cost-shape fix with no modeling surface.

— sent from bold-crane-271