[Infernal-Crack-LED]

Automated by the daily new-unit watch (~/.nikke-newunit-autopilot/run.sh): bakery-bot's sync seeded new units; nikke-sim synced them and ran the kit-autonomy gauntlet per unit (Claude-only model routing).

Units

Gauntlet results

Marian: Queen (marian-queen) — GO — faithfulness 1.0 — tier 2

• What it does: S1 fires seven 125.51% spears every 4s (interval). Each hit applies Open Wounds to the boss: Damage Taken +5.05%, permanent, non-stacking, plus the named status. Her burst cast re-applies it. S2 opens Penetrant Wish on her own cast for 10s: every 2 in-state rounds she fires a 14.06% core strike, and the spear interval drops to 2s. At Full Burst entry she gets Core Damage +24.64% if in state, or Sequential Damage +85.17% if not (ownBurstGate notCast). Burst: full reload, ATK +78.4%, Attack Damage +78.14% for 10s, and a 1346.4% nuke.
• Engine: no change. Every line uses an existing primitive.
• GAP: MG Ramp-Up Speed ▲100% has no wind-up speed modifier in the engine. Logged as engine-modeling-gaps §23, with an opt-in mgRampSpeedPct proposed after a focus recording. Only ~1s of one post-reload climb falls inside her window.
• Evidence: S2a tests 38 passed + 1 GAP skip, every counterfactual RED (S2d). S5 blind tests run 14/14 green against the override. S6 blind override structurally identical. S7 GO 1.0 (claude-fable-5-1, binding) and GO 1.0 (claude-fable-5). Same-family only (Claude).
• S2c fix: both S2b reviewers and S6 read the 2-attack count as in-state only, so I set countScope: 'gated' before landing.
• Owner spot-check, scope: the 14.06% strike is core-eligible at the MG core rate 0.95, not a guaranteed core (one reviewer read it as guaranteed, ~5% of strike damage). Open Wounds is assumed non-stacking.
• Owner spot-check, duration and trigger identity: Penetrant Wish runs 10s from her own cast (not Full Burst entry). The first spear volley fires at t=4s by engine convention, not t=0.
• Board: no recording and no graded comp includes her. regression.ts output is byte-identical with and without her override. Control comp sim total ≈1.44B (normals 500M / spears 443M / strikes 431M / nuke 67M).
• ⚠ Already broken at base 9aab2f0 (today's sync), not caused by this unit: red-hood is missing from characters.json. That crashes regression.ts (PA MiKa), board-read, kit-status --refresh and ranks:all, and with the sync data causes 7 regression drifts and ~47 vitest failures. Those failures are identical with and without this unit's changes. Expect verify.sh RED until the sync is fixed.
• Because --refresh crashes, I added her kit-status row in the --refresh entry shape and then ran --gauntlet. Unit pages and sitemap got her entries only; the sync fallout for other units is left for batch reconciliation.
• Leftover: a throwaway worktree /tmp/mq-basewt (the hook blocked forced removal). Clean it up with rm -rf /tmp/mq-basewt && git worktree prune; it is logged in docs/handoffs/autonomous-edit-queue.md.

Rapunzel: Sanctuary (rapunzel-sanctuary) — GO — faithfulness 1.0 — tier 2

• Override: each full-charge shot heals all allies (a recovery event; the HP amount isn't modelled) and adds a 220.7% hit on the boss (plain hit, can crit, not core). At every Full Burst entry, self + "Into the Fray" allies get Attack Damage ▲27.45% and ATK ▲29.17% of her ATK for 10s. Her own burst cast gives all allies ATK ▲34.84% of her ATK for 10s. Charge-speed immunity via charFixes.statImmunities. 31/31 unit tests green; simSupported: true.
• Engine primitive: a new opt-in target, alliesInStatus (statusPrefix, includeSelf), for "self and all allies in the Into the Fray state" (commit 535abad). Code review by claude-fable-5-1 returned FIX-BEFORE-MERGE; the fixes landed in a2f5bcf (a roster-wide producer warning in the validator, a ranks-page guard). scripts/regression.ts output is byte-identical with and without the engine change and the override.
• Owner action: no shipped override emits an "Into the Fray: …" status yet, so both Full-Burst-entry buffs reach only her today. For the buffs to reach Marian: Queen / Scarlet: Goddeath, marian-queen needs a battle-start selfStatus "Into the Fray: Sharpened Resolve" and "Penetrant Wish" renamed in full, and scarlet-goddeath needs "Into the Fray: Serene Mind". Exact edits are in docs/engine-modeling-gaps.md §24. I didn't make them: other units' overrides are outside this run's authority.
• Unmodelled lines:
  • Nanoparticle Injection branch (revive, +40s burst cooldown, Consumption DEF): unreachable, because nobody is incapacitated in v1.
  • Shared shield and Goddess Blessing DEF: these need Full Charge held ≥1s, which the sim never does. Same precedent as rapunzel-pure-grace, guilty-mighty-bunny and sin-swift-bunny.
• Owner spot-check:
  • SCOPE: Into the Fray allies only (dormant until the edits above).
  • DURATION: all grants 10s; the immunity is whole-fight here vs from her first full-charge shot in-game.
  • TRIGGER-IDENTITY: the 27.45 / 29.17 buffs fire at every Full Burst, including ones she sat out; the 34.84 buff fires on her own cast; check whether auto ever holds a charge.
• Reviewers: S2b fable-5 and fable-5-1 agreed with the driver's spec on every line. S6 matched on every damage line. Both S5 and S6 put the held-charge lines on every full-charge shot, which both judges rejected. S5's blind tests (after a one-line harness-wiring fix) passed 7/10 against the override: 2 RED are that held-charge proxy, 1 is a noFb arm the engine ignores by design.
• Board: no recording and no graded comp, so no A/B numbers. The 7 regression snapshot drifts and the 5 failing ranks tests already fail at this branch's base: the sync dropped red-hood.
• Skipped / leftovers:
  • Full verify.sh not run (left to the wrapper).
  • Two tagged stash entries (rapsanct-*) left because git stash drop is hook-blocked; they duplicate committed work and are noted in docs/handoffs/autonomous-edit-queue.md.
  • The untracked scarlet-goddeath portraits were left alone.

Scarlet: Goddeath (scarlet-goddeath) — NO SUMMARY (session died or hit a ceiling) — see /Users/maxwellsutton/Library/Logs/nikke-newunit-watch/run-20261029T060005

Roster test pins

roster test pins — 2026-10-29 sync: NOTHING COMMITTED (upstream data regression, not pinnable)

• The sync (9aab2f0) is broken upstream, so I updated no pins. It added marian-queen / rapunzel-sanctuary / scarlet-goddeath, but it also dropped red-hood (her override is still in src/skills/overrides/red-hood.json; she's also gone from gauge-per-shot.json, archetype-tags.json and tsareena-build.json) and set rl3 + releaseDate to null on about 196 of 205 characters (rl3: 0 nulls before, 196 after; all 21 Treasure dates are gone). This looks like bakery-bot's Synergy attributes/rows going missing. Pinning it would write the bad data into the tests.
• Typecheck is green. scripts/tests without units/: 43 failed / 1407 passed in 15 files. The same 5 drift suites all pass at HEAD~1, so the sync caused these failures:
  • red-hood gone: census-kit-numbers (2), census-unmodeled-entries (2), convert-excess, stage-targeting, ranks/b1b2dps (4), battery/multihit-crediting, and probably gauge-source-census (its fixture names red-hood).
  • releaseDate nulled: data/release-dates (22), share/new-favorite-items (4).
  • Gauge/frame drift: element-control and charge-counter-gates (2). Most likely red-hood or rl3 going missing; I didn't isolate which.
  • New units: ranks/sustain (rapunzel-sanctuary is tagged healer and has no SUSTAIN_TABLE entry, so it needs modeling, not a pin). share/sitemap-drift and serve-api/serve-headers (sitemap 404) changed with the roster too, but I left the sitemap alone because regenerating it would lock in red-hood's removal.
• Before any PR: fix the upstream DB (or sync.ts's matching) and re-run npm run sync, or revert the sync's red-hood/rl3/releaseDate losses. Don't push this branch as it stands.
• Untracked new-unit portraits in web/public/img/portraits/ were left as they were (not mine).

Gate

bash scripts/verify.sh: RED (log on the owner's Mac: /Users/maxwellsutton/Library/Logs/nikke-newunit-watch/run-20261029T060005/verify.log)

Same-family provenance: every blind/judge role ran on a Claude model (no Kimi/Qwen access) — a GO is weaker evidence than a cross-family GO; spot-check the scope / duration / trigger-identity lines.

🤖 Generated with Claude Code