Overnight swarm report — 2026-07-15 morning handoff

Overnight swarm report — 2026-07-15 morning handoff

Run per docs/overnight-swarm-runbook-2026-07-14.md, launched ~22:40. Honest caveat first: the Mac slept from ~23:30 to ~07:10 despite Amphetamine + Low Power Mode, so the “10-hour” swarm compressed into two active windows (~22:40–23:30 and ~07:10–08:15). Within those windows all four dispatched workers completed end-to-end. A concurrently-running session also landed soak/coverage work early in the night (its own commits, coordinated via the punch-list claim convention — which held up, with one benign shared-file sweep noted below).

Completed (all pushed to main)

1. §2 RC-BLOCKER FIXED — Wyrm shared-HP AoE multiplication (dc3e739, W1). Root cause: damage_enemy mirrored each segment hit into the worm’s shared pool with no per-event dedup, so an AoE covering N segments dealt N× damage. Fix: ElementalSystem. begin/end_aoe_group() — once-per-pool-per-source-event, all 17 multi-target loops bracketed. 7 new tests (burst/zone/nova each failed pre-fix); suite green; determinism baselines unchanged, no re-pin; soak clean (4 runs, 0 invariants).
2. §2 solo SOFTLOCK flags — verdict: false alarm (39d0b16, W3, evidence-based). Seed-3000 instrumented replay: AutoPlayer engages fine; the “stall” is a legitimate 3.6-min solo boss duel (~6.6k HP at ~31 HP/s) tripping the 3-min zero-kill heuristic; co-op’s 2× DPS stays under the window (hence 0/6 co-op flags). No fix applied — correctly, since both candidate fixes are design decisions (below).
3. §2 ghost flag — CLAUDE.md no longer references V01_LOCK_NETWORKED_MP as if on main (221ea4f, W2).
4. §4 batch (W2): dead _joy_lbl/_update_joy_debug_label() removed from main.gd with full ritual incl. foreground ui-sweep 9/9 ok (3a52f42); CLAUDE.md RNG-streams 2→4 corrected; SOUND_MAP.md +12 honest rows (59 DONE · 13 GENERIC · 16 SILENT of 88); docs-site gitignore item found already-done 07-05, stale entries closed (221ea4f).
5. §2 Obsidian cruiser — 3 real sprite candidates ready (c72ecbc/ce3e282, W4): tools/ship_preview/candidates-obsidian/ — obsidian_bastion (armored wedge, deep violet), obsidian_ram (twin-prow + glowing ram trim), obsidian_monolith (hammerhead slab, indigo), plus obsidian_lineup.png against the cobalt frigate at the in-game 1.5× ratio. Baked with the shipped-hull recipe via a new reusable harness (tools/ship_preview/obsidian_candidates.*).

Decisions / picks awaiting Chris

• Cruiser sprite pick (§2, still ⏳ on the list): choose bastion/ram/monolith. Two wiring notes on the list: TV-legibility of the dark flanks; PlayerRenderer baked-sprite scale is tier-based, not radius-based — the 1.5× visual size needs a wiring decision.
• Sim._spawn_boss_adds is orphaned dead code (new §2 item, W3’s real discovery): survival PHASE_BOSS still runs the old clean-arena suppression its own comment says was replaced. Wiring it in draws from rng → determinism re-pin + real pacing change (would also shorten the solo boss-duel flatlines) — deliberate call, not an overnight fix.
• Balance evidence: first soak data is on the §2 balance item (56 runs + coverage sweeps, from the concurrent session). Notable: Aim weapon ≈60% of all soak damage.

New findings filed on the punch list (per the standing rule)

• §4: Blaster starter + manual Aim weapon fire completely silently — the game’s top damage dealers emit no fire fx_event at all (W2).
• §4: per-segment status DoT — burn applied to N wyrm segments drains the shared pool N× via apply_status_and_decay; separate mechanism from the fixed AoE-event bug, needs a design call (W1).
• §4: dash-into-solid-hazard damage bypasses dev_invuln (deliberately bypasses is_invulnerable()) — killed an “invuln” soak run in void; minimal one-condition fix proposed (W3).

State of main at handoff (HEAD 39d0b16, == origin/main)

• Full GUT suite: 232/232 scripts, 2059 passing, 0 failing (4 pre-existing risky/pending); check-test-count.sh OK; boot smoke clean (0 SCRIPT ERROR) — re-run at 07:45 on final HEAD.
• Determinism: both pinned baselines pass unchanged all night (zero re-pins).
• UI sweep: 9/9 ok vs blessed baselines (last full run on 3a52f42, the final code change).
• Claimed-but-unfinished: none (the one open ⏳ is the cruiser pick, deliberately Chris’s).

Process notes for the next overnight run

• The Mac still slept. Amphetamine was running but the machine suspended ~23:30–07:10 — worth checking the Amphetamine session settings (“Allow system sleep when display is closed” must be UNCHECKED, and it’s a Session Default — a session started before setting it doesn’t inherit) or using sudo pmset disablesleep 1 next time. Workers resumed cleanly on wake and finished — the claim convention meant nothing was lost, only time.
• Shared-checkout concurrency held, with one benign race class seen twice: a worker committing the shared punch list swept another worker’s in-flight claim edit into its commit (content verified correct both times). If overnight parallelism grows, consider per-worker claim commits BEFORE any other file edits, or worktree isolation for punch-list edits.
• .git/index.lock contention was frequent but the wait-and-retry protocol resolved every case; no locks deleted, nothing lost.