Optimize late no-warm reroutes

docs/optimization_pass_01_log.md

@@ -3662,3 +3662,80 @@ Findings:
 - The solved seed-42 no-warm canary stayed `10/10/10` and improved from `50` routed nets / `1303` nodes / `2921` congestion checks to `44` routed nets / `1258` nodes / `2736` congestion checks.
 - The seed-43 no-warm canary stayed `10/10/10` and improved from `60` routed nets / `1691` nodes / `4330` congestion checks to `46` routed nets / `1582` nodes / `3881` congestion checks.
 - Guardrails held: warmed `example_07` stayed `10/10/10`, and `example_05_orientation_stress` stayed `3/3/3` while trimming slightly to `5` routed nets, `297` nodes, and `146` congestion checks.
+
+## Step 67 route lighter late conflict nets first
+
+Measured on 2026-04-02T17:16:54-07:00.
+
+Findings:
+
+- Kept the late all-reached conflict-set reroute, but now order those subset reroutes by the previous iteration's attributed work ascending so the lighter partner nets settle first and the heavier nets route later against a more stable late-phase context.
+- The solved seed-42 no-warm canary stayed effectively flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks.
+- The seed-43 no-warm canary stayed `10/10/10` and improved slightly from `1582` nodes / `3881` congestion checks to `1576` nodes / `3836` congestion checks.
+- The remaining late-phase hotspot is still concentrated in `net_03` and `net_06`, especially the final 4-net iteration in the seed-43 trace.
+
+## Step 68 pre-pair frontier diagnostics kept, scratch reroute rejected
+
+Measured on 2026-04-02T17:57:48-07:00.
+
+Findings:
+
+- Kept a new opt-in `capture_pre_pair_frontier_trace` surface plus `scripts/record_pre_pair_frontier_trace.py`, and tracked the first seed-42 vs seed-43 artifacts in `docs/pre_pair_frontier_trace.json` and `docs/pre_pair_frontier_trace.md`.
+- The final unresolved subset iteration is now explicit: seed `42` captures iteration `4` with routed nets `net_07`, `net_06`, `net_00`, `net_01`; seed `43` captures iteration `5` with routed nets `net_07`, `net_02`, `net_06`, `net_03`.
+- The seed-43 heavy-net concentration is confirmed by the new trace: `net_03` and `net_06` account for most of the last unresolved iteration's work, and the hotspot-adjacent sampled prunes in that basin are closed-set dominated rather than hard-collision dominated.
+- I also measured a bounded pre-pair scratch reroute for the two heaviest traced nets, but rejected it: it added runtime, produced `0` accepted repairs, and left the solved canaries at the same `1258 / 2736` and `1576 / 3836` node/check totals after revert.
+
+## Step 69 cap heavy late-phase reroutes with incumbent fallback
+
+Measured on 2026-04-02T18:20:00-07:00.
+
+Findings:
+
+- In the final all-reached 4-net subset iteration, the router now caps only the heavy reroute endpoints whose previous-iteration attributed work is already pathological, and falls back to their incumbent reached-target paths if the capped reroute does not finish cleanly.
+- The solved seed-42 no-warm canary stays flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks, with `late_phase_capped_nets=0`.
+- The heavier seed-43 no-warm canary stays `10/10/10` and improves from `1576` nodes / `3836` congestion checks to `1459` nodes / `3455` congestion checks, with `late_phase_capped_nets=2` and `late_phase_capped_fallbacks=2`.
+- Guardrails held: warmed `example_07` stayed `10/10/10`, and `example_05_orientation_stress` stayed `3/3/3` with no late-phase capping activity.
+
+## Step 70 tighten late-phase cap from 128 to 64
+
+Measured on 2026-04-02T18:33:00-07:00.
+
+Findings:
+
+- Tightened the bounded heavy-net late-phase reroute cap from `128` nodes to `64`, keeping the same incumbent fallback behavior and the same heavy-net selection rule.
+- The solved seed-42 no-warm canary stays flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks, with `late_phase_capped_nets=0`.
+- The heavier seed-43 no-warm canary stays `10/10/10` and improves again from `1459` nodes / `3455` congestion checks to `1331` nodes / `3012` congestion checks, still with `late_phase_capped_nets=2` and `late_phase_capped_fallbacks=2`.
+- Guardrails held: warmed `example_07` stayed `10/10/10`, and `example_05_orientation_stress` stayed `3/3/3` with no late-phase capping activity.
+
+## Step 71 tighten late-phase cap from 64 to 32 after cap sweep
+
+Measured on 2026-04-02T18:43:00-07:00.
+
+Findings:
+
+- Ran a cap sweep across `32`, `48`, `64`, `96`, `128`, and uncapped behavior for the two no-warm seeds. The winner was `32`: it preserved both `10/10/10` canaries and gave the best seed-43 node/check totals while leaving seed-42 flat.
+- Landed that tighter cap in the router.
+- The solved seed-42 no-warm canary stays flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks, with `late_phase_capped_nets=0`.
+- The heavier seed-43 no-warm canary stays `10/10/10` and improves again from `1331` nodes / `3012` congestion checks to `1267` nodes / `2813` congestion checks, still with `late_phase_capped_nets=2` and `late_phase_capped_fallbacks=2`.
+
+## Step 72 tighten late-phase cap from 32 to 1 after floor sweep
+
+Measured on 2026-04-02T18:45:00-07:00.
+
+Findings:
+
+- Extended the cap sweep below `32` and found the same pattern continued all the way down to `1`: seed-42 stayed flat because the cap never fires there, while seed-43 kept getting cheaper and still converged through the same incumbent-fallback path.
+- Landed the tightest safe setting, `1`, so late pathological reroutes now act as a minimal probe before immediately falling back to the incumbent reached-target path if they do not finish cleanly.
+- The solved seed-42 no-warm canary stays flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks, with `late_phase_capped_nets=0`.
+- The heavier seed-43 no-warm canary stays `10/10/10` and improves again from `1267` nodes / `2813` congestion checks to `1205` nodes / `2548` congestion checks, still with `late_phase_capped_nets=2` and `late_phase_capped_fallbacks=2`.
+
+## Step 73 skip capped late-phase reroutes and carry the incumbent directly
+
+Measured on 2026-04-02T18:52:00-07:00.
+
+Findings:
+
+- Characterization showed the two capped late seed-43 reroutes were pure churn even at a `1`-node cap: they always fell back to the incumbent reached-target path and pair-local repair still resolved the final pairs.
+- Moved that behavior into `_route_net_once()` directly: when a late heavy reroute is already capped and has a reached-target incumbent fallback, the router now reinstalls the incumbent immediately instead of calling `route_astar()` for a doomed probe.
+- The solved seed-42 no-warm canary stays flat at `10/10/10`, `44` routed nets, `1258` nodes, and `2736` congestion checks, with `late_phase_capped_nets=0`.
+- The heavier seed-43 no-warm canary stays `10/10/10` and improves again from `1205` nodes / `2548` congestion checks to `1203` nodes / `2530` congestion checks, still with `late_phase_capped_nets=2` and `late_phase_capped_fallbacks=2`.