v0.27.0 — summary quality

H1 step 4: summary quality. rc1 about 2026-10-17. The fixes below are draft PRs, held until the v0.26.0 GA. After the GA they are rebased and merged in dependency order. #905, #910, #912 and #764 go first because they share the leaf call site.

Scope:

• Summariser input (#611 -> #905, decided): the leaf's message texts share the leaf token budget instead of a fixed per-message clip. Synthetic measurement: middle-placed facts reaching leaf summaries 0/72 today vs 42/72 with the fix. Also closes #899.
• Prompt default (#660): resolved by the flip rule. Prompt v1 stays the default. v2 failed the rule in its first run, and the v1 control (#646) cancelled the re-run. A v2 default returns only if a later v2 run passes the same rule: v2 >= v1 on every paired axis, better on at least one, latency within the bars, and a level-3 rate within band of zero. The change stays an owner decision.
• Prompt v2 for opt-in users: #646 -> #912. #912 merges only after its bench gate: at least 3 interleaved v1 vs v2 runs on the rebased head, reporting the median and range. It passes when the median meets the v1-control bars (accepted rate >= v1, condensation p90 <= v1 + 10 s, max <= 120 s). Two runs on one head differed by more than 0.10, so one run is not enough.
• Survival fits and turn continuity: #900 -> #918, #798 -> #920, #916 -> #917 (rollback advice #919 -> #928), #909 -> #910, #751 -> #764.
• Tool-loop defects found by the new continuity cell (#914): #921 -> #925, #922 -> #933, #923 -> #924. The hold length moved to #926 (v0.28.0).
• Outside reports: #904 -> #935, #936 (#930) and #934 (#931); #903 -> #927; #852 -> #929.
• Small follow-ups: #897 -> #913, #906 -> #915, #613 -> #746, #438 and #240 -> #747. Cue lines (#663) if cheap.
• Instruments: #914 (continuity check after each compaction), #907 (Track S run kit), #911 (CI host id rename).
  Off-theme items moved to v0.28.0 or Backlog on 2026-10-06.

GA accepts when:
(1) No LCM-X system-level axis is below the v0.26.0 row. The rule is Track S paired per-fact exact McNemar on identical v3 material, alpha 0.05, minimum effect 5 points, with runs added when the run1/run2 spread exceeds 0.10. QA rows use their own paired rule. The v0.26.0 row is run right after the v0.26.0 GA.
(2) The theme gain is present: middle-placed facts > 0 after #611, and constraint continuity >= the floor declared from the v0.26.0 baseline.
(3) Compaction wall time on the >= 24 h reference-agent soak: p90 <= 60 s and max <= 120 s. Gauntlet cells: max <= 145 s.
(4) Recall (#898): 0 per-question lcm_recall flips vs v0.26.0 with embeddings off. Otherwise each flip must be attributed and classified intended, within the same floor as v0.26.0.
(5) 0 respin-bar defects.
Comparisons with Codex native, lossless-claw, plain Hermes and Claude Code are reported as H1 exit evidence, not per-release gates (#664). The continuity suite (#659) is recorded in this release and gates from v0.28.0.