Audit tools are test code (bug-finders), not scripts. Move to tests/audit/.
scripts/ now only replay-combat (live demo tool).
scratch/ = gitignored throwaway. Repro scripts, exploration, debug.
Update DEVELOPMENT.md + scripts/README to match new layout.
Audit bug: pause fired turn%12, no resume in same iter. nextTurn then
called on paused encounter → threw 'Encounter not running'. Throw is
correct feature behavior (nextTurn refuses when paused); audit misuse.
Fix: togglePause twice (pause+resume) in one iteration, plus guard
'advance-while-paused' check before nextTurn call.
Result: 6 audit artifacts → 0 violations / 100 rounds.
Confirms BUG-1 resolved as side effect of BUG-2 dup-id fix.
Replay verify: 10 rounds, 103 turns, no skip/dupe.
TODO: BUG-1 + BUG-2 marked RESOLVED/FIXED.
Root cause: addParticipant appended participant to participants[] without
checking id uniqueness. Two participants with same id in array. On
togglePause resume, turnOrderIds rebuilt via sort → dup id appears twice.
nextTurn then stuck repeating that id (rotation breaks).
This was the enabling step for BUG-1's full corruption (audit chain):
pause blocks advance → totalTurns frozen → addParticipant re-adds
same r${totalTurns} id → resume dup → nextTurn stuck.
Fix: throw on duplicate id in addParticipant. Caller must use fresh id
(crypto.randomUUID in App, replay already does).
Evidence:
- Test: 'addParticipant rejects duplicate id' (was test.skip, now live).
- Pre-fix: 1 RED (Received function did not throw).
- Post-fix: 50 green (shared), 23 green (server), 62 green (FE).
- Reachability in normal app: low (App uses crypto.randomUUID) but no
guard existed before. Defensive + unblocks BUG-1 isolation.
No other behavior changed.
Runs pure turn.js combat, audits per round:
- rotation integrity, HP bounds, isActive consistency, turnOrder dup,
currentTurn valid, deathSave range, removeParticipant orphans,
conditions, undo support
100-round run: 128 violations all BUG-1/BUG-2 family (4 symptoms).
Clean: HP, isActive, deathSave, conditions, removal.
Exploratory (Math.random), not unit test. Unit tests lock known bugs.
Move all test files out of source dirs into per-workspace tests/:
- shared/tests/ (3 unit test files)
- server/tests/ (1 integration test)
- src/tests/ (8 characterization + scenario tests + testHelpers)
Fix all relative import paths (App, storage, __mocks__, testHelpers).
Fix jest.config testMatch globs in shared/ and server/ (rootDir +
<rootDir>/tests pattern).
Delete scripts/repro-pause-bug.js (debug scratch, superseded by
turn.pause-add.test.js).
Keep scripts/replay-combat.js + scripts/audit-rotation.js as manual
demo/exploratory tools (NOT unit tests, not deterministic).
No logic changes. All green: shared 49 + 1 validated RED, server 23,
FE 62. Scenario test unchanged (240s timeout, pre-existing slow).
- turn.round-rotation.test.js: 7 tests, full round visits each active
participant once (pure nextTurn clean). Green.
- turn.characterization.test.js: RED 'addParticipant rejects duplicate id'.
Validates current behavior allows dup ids (self-inflicted in audit via
loop spin-while-paused re-adding same id; unreachable in app via
crypto.randomUUID, but documents gap).
- audit-rotation.js: pure turn.js simulation of replay op sequence.
Detects rotation violations (skip/dupe per round). Pause disabled = 0
violations across 100 rounds. Pause enabled = 56-77 violations starting
round 20. Pinpoints addParticipant+pause interaction.
- repro-pause-bug.js: minimal repro scripts.
- replay-combat.js: rewritten for real rounds (full initiative cycles),
visible damage each turn, all conditions, toggleActive, remove,
reinforce, edit, pause/resume, reorder, endEncounter. HP bumped for
100-round sustain + revive dead each round.
No feature code changed.
- ROUNDS now = full initiative cycles (not turns). Each round advances
initiative until round counter ticks (all participants act).
- Visible damage: current actor hits random living target for 3-10 dmg.
Player view sees HP bars change live.
- Default delay 200ms (was 800ms).
- Reproduces M4 skip bug: rounds shrink as participants die (8→7→2→1).
- Label accuracy: 'turn N (round X)'.
Root cause (HAR-diagnosed): replay script wrote firebase-prefixed paths via
raw REST, bypassing adapter norm(). Two path roots coexisted in db:
bare 'campaigns/X' (adapter writes, from App)
prefixed 'artifacts/.../campaigns/X' (replay raw writes)
Adapter read bare, missed prefixed. UI showed stale test1 (legit manual UI
write, not wiped) but replay campaigns invisible.
A. replay-combat.js: use createWsStorage adapter instead of raw fetch. Same
contract boundary as App. norm() runs on all paths. Can't drift.
Mirror App.js getPath locally for path construction.
B. contract.js: 4 new identity tests (setDoc prefixed -> getCollection bare,
setDoc prefixed -> getDoc bare, setDoc prefixed -> getDoc prefixed,
setDoc bare -> getCollection prefixed). Run against every impl (memory,
ws). memory.js lacked norm() -> RED first, now GREEN after adding norm.
C. db moved out of /tmp to ./data/tracker.sqlite (gitignored). Never tmp.
Tests: 124 green (39 shared + 23 ws-contract + 62 FE).
scripts/replay-combat.js: drives full combat via live backend REST, computes
turns through shared/turn.js. Player display (subscribed via WS) live-updates.
Usage: node scripts/replay-combat.js [rounds] [delayMs]
TODO.md: tracks M4 work.
- Dead participants must NOT be skipped (still occupy initiative slot,
death saves resolve on their turn). Saw in game Saturday.
- JUMP_TURN_TO manual turn override.
Drives same UI buttons a DM clicks. Exercises:
- campaign + character roster (3 PCs)
- monsters, NPC, add-all, hidden hp toggle
- start combat, 100 rounds of nextTurn
- damage/heal, conditions (stunned), toggle-active
- edit initiative, edit name, remove participant
- pause/resume with reinforcement mid-combat
- damage-to-0 + death saves (x3) + revive
- end combat + confirm
Harness: each phase wrapped in recordAsync try/catch. Failures collected,
reported at end, do NOT abort run. Found no app bugs; selector bugs in harness
fixed (char form ids vs placeholders, encounter row scope via Init: marker,
conditions panel async render, dead-participant damage skip as expected game
state, end-encounter modal title 'End Encounter?' not 'End Combat').
289/289 phases ok. 58 other frontend + 39 shared + 19 ws-contract = 116 total green.
DisplayView missed in original M2 refactor — raw onSnapshot(doc(db,path))
survived. In ws/memory mode db is a stub sentinel, so raw SDK calls crash
('Expected first argument to collection() to be a CollectionReference...').
Reported by human testing player display after start combat.
TDD:
1. RED: DisplayView.characterization.test.js asserts adapter.subscribeDoc
called for campaign + encounter + activeDisplay paths. Adapter recorder
(getAdapterCalls/resetAdapterCalls in firebase.js) instruments subscribe
calls — catches raw-SDK bypass that firebase mock alone cannot (mock db
satisfies raw onSnapshot, hiding the bug).
2. Fix: 2 raw onSnapshot sites in DisplayView -> storage.subscribeDoc.
3. GREEN: 2 new tests pass, 116 total green.
Audit confirmed DisplayView was the ONLY remaining raw SDK site in App.js.
ws/memory storage mode never set the module-level `db` variable. 24 handlers
guarded with `if (!db) return` early-exited, silently dropping all writes
(create campaign, add encounter, participant CRUD, combat, logs).
db stays a truthy sentinel object { __localStub: true } in non-firebase mode.
All real reads/writes route through storage.*; db only used by guards.
56 frontend tests green. Verified via headed browser: create campaign flow
works end-to-end (modal closes, campaign appears via WS realtime push).
Firestore rejects writes containing undefined values. The pause/resume
and end-combat undo snapshots read encounter fields that may not yet
exist in Firestore, so add ?? false / ?? null / ?? 0 fallbacks to
match the pattern already used in the start-combat undo path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a participant was activated mid-combat, computeTurnOrderAfterAddition
appended them to the end of turnOrderIds. The visual display sorted by
initiative (putting them at the top), but the turn pointer followed the
append order, making it look like the top-initiative participants were
skipped when the round wrapped.
Fix: at the round boundary in handleNextTurn, rebuild turnOrderIds from
all active participants sorted by initiative. Mid-round additions go last
in round 1 (standard D&D ruling), then slot into proper initiative order
from round 2 onwards. Also adds turnOrderIds to the next-turn undo snapshot.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Each logged action now stores a Firestore snapshot of the affected
encounter state. The /logs page shows an ↩ Undo button on any entry
with undo data; clicking it restores the encounter to its pre-action
state and marks the entry as rolled back (greyed out, strikethrough).
Covered actions: damage/heal, condition toggle, activate/deactivate,
add/remove participant, next turn, start/pause/resume/end combat.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
david raistrick
robert