Match cache survives releases; v2.0.1 round-robin results

docs/STRATEGY_DEVELOPMENT.md

@@ -147,7 +147,8 @@ reaches `main` by pull request:
    `git add` new files before `pixi run lint`: it checks tracked files only. Before waiting on
    a `--reuse` round-robin, read the "N reused ... M to play" line it prints: M much larger
    than your own matches means the cache is stale (shared code changed since it was filled).
-   Stop it and ask for a full refresh on `main` rather than paying for one on your branch.
+   Let it run anyway (a full round-robin is about 10 minutes on 15 workers): the other
+   strategies' matches run `main`'s code, so they refill the cache for everyone.
 5. Open a draft pull request into `main` with the round-robin result, and the bench result if you
    ran one. An agent never merges: a person reviews and merges.
 

docs/strategies.md

@@ -55,36 +55,36 @@ only from `--both-sides`, side/kickoff cells, or `--fuzz-restarts`.
 
 ## Latest results
 
-**Smoke round-robin, full catalog** (2026-10-01, `e83a7466`, 22 configs, 231 matches, 65s,
-`--strict`, 0 stalls, `replays/tournament_20261001_094103/summary.json`). Points are 3 a win,
-1 a draw. 65s matches are mostly draws (152 of 231 here), so the middle of the table is close
-to noise; two round-robins of near-identical code rank strategies with Spearman +0.97
-(`docs/STRATEGY_DEVELOPMENT.md`, A/B on the scenario bank).
+**Smoke round-robin, full catalog** (2026-10-03, v2.0.1, 22 configs, 231 matches, 65s,
+`--strict`, 0 stalls, about 9 min wall on 15 workers, `replays/tournament_20261003_092151/summary.json`).
+Points are 3 a win, 1 a draw. 65s matches are mostly draws (129 of 231 here), so the middle of
+the table is close to noise; two round-robins of near-identical code rank strategies with Spearman
++0.97 (`docs/STRATEGY_DEVELOPMENT.md`, A/B on the scenario bank).
 
 | Strategy | W-D-L | GF-GA | Points per match |
 |---|---|---|---|
-| `split_shape` | 8-13-0 | 11-1 | 1.76 |
-| `clear_danger` | 6-15-0 | 7-0 | 1.57 |
-| `overload_press` | 7-11-3 | 9-3 | 1.52 |
-| `counter_flow` | 5-16-0 | 6-0 | 1.48 |
-| `score_aware_counter_flow` | 5-15-1 | 5-1 | 1.43 |
-| `press_and_pass` | 5-14-2 | 5-2 | 1.38 |
-| `high_press` | 3-18-0 | 3-0 | 1.29 |
-| `low_block` | 4-15-2 | 4-2 | 1.29 |
-| `give_and_go_solo` | 5-12-4 | 6-7 | 1.29 |
-| `clear_press_plus` | 3-17-1 | 3-1 | 1.24 |
-| `counter_press` | 4-14-3 | 5-4 | 1.24 |
-| `press_trigger_flow` | 4-14-3 | 5-4 | 1.24 |
-| `high_line_zone` | 4-13-4 | 4-4 | 1.19 |
-| `tiki_taka_plus` | 3-15-3 | 3-3 | 1.14 |
-| `tiki_taka` | 2-17-2 | 2-2 | 1.10 |
-| `decoy_and_overload` | 3-13-5 | 3-6 | 1.05 |
-| `three_slot` | 1-17-3 | 1-3 | 0.95 |
-| `switch_of_play` | 3-9-9 | 3-9 | 0.86 |
-| `overload_flow` | 1-13-7 | 2-9 | 0.76 |
-| `zone_fluid` | 1-12-8 | 2-10 | 0.71 |
-| `shadow_switch` | 2-8-11 | 4-13 | 0.67 |
-| `score_aware_zone_flow` | 0-13-8 | 0-9 | 0.62 |
+| `counter_flow` | 9-12-0 | 9-0 | 1.86 |
+| `clear_danger` | 10-9-2 | 10-2 | 1.86 |
+| `split_shape` | 10-8-3 | 12-3 | 1.81 |
+| `give_and_go_solo` | 8-11-2 | 9-3 | 1.67 |
+| `score_aware_counter_flow` | 6-14-1 | 6-1 | 1.52 |
+| `decoy_and_overload` | 7-9-5 | 8-5 | 1.43 |
+| `low_block` | 5-14-2 | 5-2 | 1.38 |
+| `overload_press` | 4-16-1 | 5-1 | 1.33 |
+| `press_and_pass` | 4-16-1 | 4-1 | 1.33 |
+| `press_trigger_flow` | 6-10-5 | 6-5 | 1.33 |
+| `clear_press_plus` | 4-15-2 | 4-2 | 1.29 |
+| `high_press` | 4-14-3 | 4-3 | 1.24 |
+| `tiki_taka_plus` | 3-16-2 | 3-2 | 1.19 |
+| `counter_press` | 4-11-6 | 4-6 | 1.10 |
+| `high_line_zone` | 4-11-6 | 4-7 | 1.10 |
+| `tiki_taka` | 2-16-3 | 2-3 | 1.05 |
+| `overload_flow` | 4-8-9 | 5-11 | 0.95 |
+| `three_slot` | 2-12-7 | 2-7 | 0.86 |
+| `score_aware_zone_flow` | 3-7-11 | 4-12 | 0.76 |
+| `switch_of_play` | 2-10-9 | 2-10 | 0.76 |
+| `shadow_switch` | 0-13-8 | 2-11 | 0.62 |
+| `zone_fluid` | 1-6-14 | 1-14 | 0.43 |
 
 The full-match tables below predate the planner, referee and tactic fixes since 2026-09-02.
 Treat them as rough ordering, not current truth.

utama_core/replay/fingerprint.py

@@ -36,6 +36,7 @@
 import hashlib
 import json
 import os
+import re
 import sys
 from dataclasses import dataclass
 from functools import cached_property
@@ -60,6 +61,10 @@
 # Files next to a module that are not data it reads.
 _NOT_DATA_SUFFIXES = (".py", ".pyc", ".md")
 
+# The release workflow rewrites `utama_core/__init__.py`'s `__version__` on every merge into main.
+# Nothing reads it in a match, and hashing it would invalidate every stored result on each merge.
+_VERSION_LINE = re.compile(rb"""^__version__\s*=\s*["'][^"'\n]*["']\s*$""", re.MULTILINE)
+
 # Method names that mutate their receiver; called at import time, they reach outside the module.
 _MUTATORS = frozenset(
     {
@@ -195,7 +200,8 @@ def file_digests(self, names: Iterable[str]) -> list[str]:
         for name in names:
             mod = self.module(name)
             if mod is not None and mod.path is not None:
-                rows.append(f"{mod.path.relative_to(self.root).as_posix()}  {_sha(mod.source)}")
+                code = _VERSION_LINE.sub(b"", mod.source)
+                rows.append(f"{mod.path.relative_to(self.root).as_posix()}  {_sha(code)}")
         return sorted(rows)
 
     def data_files(self, names: Iterable[str]) -> list[str]:

utama_core/tests/replay/test_fingerprint.py

@@ -85,7 +85,7 @@ def build_c_kernel_strategy(ids):
 FILES = {
     "tools/tournament/__init__.py": "",
     "tools/tournament/tournament_lib.py": "from utama_core.run.runner import run\nfrom utama_core.strategy import kernel_strategy\n",
-    "utama_core/__init__.py": "",
+    "utama_core/__init__.py": '__version__ = "1.0.0"\n',
     "utama_core/run/__init__.py": "",
     "utama_core/run/runner.py": "from ..planning import planner\nfrom utama_core.profiles import loader\n\n\ndef run():\n"
     "    return planner.plan()\n",
@@ -253,6 +253,16 @@ def test_fingerprints_do_not_depend_on_where_the_checkout_is(repo, tmp_path_fact
     assert _fingerprints(other) == _fingerprints(repo)
 
 
+def test_a_release_version_bump_changes_nothing(repo):
+    # The release workflow rewrites `__version__` on every merge into main; hashing it emptied
+    # the whole match cache after each merge.
+    rel = "utama_core/__init__.py"
+    assert _changed_after(repo, rel, '__version__ = "1.0.0"', '__version__ = "2.0.1"') == set()
+    assert _changed_after(repo, rel, '__version__ = "2.0.1"', '__version__ = "2.0.1"\nDEBUG = True') == {"base"} | set(
+        CONFIGS
+    )
+
+
 def test_a_relative_import_resolves(repo):
     g = CodeGraph(repo)
     assert "utama_core.planning.planner" in g.closure(["utama_core.run.runner"])