DM1 V1 Bug Audit — Firestaff v2.7.12 → v2.7.13

Systematic audit of the Firestaff DM1 V1 runtime against the ReDMCSB decompilation as ground truth.

Auditor: Opus 4.6 Firestaff Worker Date: 2026-06-13 Baseline: Firestaff commit f99587c35 (HEAD of main, post-v2.7.12) ReDMCSB: WIP20210206, Toolchains/Common/Source/ Build dir: /Volumes/Extern-disk/legacy-workspace-work/firestaff/build Phase A probe: 23/23 PASS DM1 test suite: 556/593 PASS (37 failures — see analysis below)

Historical baseline: the counts and host paths above describe the 2026-06-13 audit and must not be read as the current result. On 2026-08-31, the current out-of-tree build-dm1-csb-native completed all 84 tests carrying the dm1 label against the supplied real-data root without a failure record. That is regression evidence, not a claim of full-campaign or pixel parity. The current source-review status and remaining capture requirements live in docs/parity/REDMCSB_DM1_CSB_AUDIT.md.

Test Failure Analysis

Of the 37 DM1 V1 test failures:

• 12 failures: Python verification scripts reference build/test_dm1_v1_viewport_3d_pc34_compat via hardcoded path ROOT / "build" which points to the repo root, not the external build directory. The binary exists and passes in the actual build dir. These are test infrastructure issues, not code bugs.
• 2 failures: Movement tests (pass552, pass580) — same hardcoded path issue for test_dm1_v1_movement_timing_pc34_compat.
• 5 failures: Capture route/transcript tests (pass504, pass510, pass623, pass625, pass626) — need original game data files not present on the build host.
• 3 failures: Viewport occlusion chain failures (pass361, pass362, pass434) — chained gate dependencies; pass434 is the root cause (viewport crop readiness gate).
• 5 failures: Viewport wall/door/field source-lock tests (pass499, pass565_d0c, pass608, pass650, pass577) — depend on missing symbols or chained prerequisites.
• 2 failures: pass519, pass570 — viewport field/front order source locks with chained dependencies.
• 1 failure: pass560 — mirrored door front, known pre-existing failure.
• 2 failures: Runtime timeouts (dm1_v1_hall_walkaround_runtime, dm1_v1_champion_mirror_candidate_panel_runtime) — require real game data; 10s timeout insufficient.
• 3 failures: pass561, pass562, pass563 — viewport wall source locks, missing binary.
• 2 failures: Additional gate chain failures.

Bug List

BUG-101 FIXED — Armor Defense Uses F0321 Wound Defense

• Severity: Resolved runtime
• Category: Mechanics
• Description: The former skill-level approximation has been replaced by the shared F0321/F0313 compatibility route.
• ReDMCSB Reference: CHAMPION.C F0321:1838-1900, F0313_CHAMPION_GetWoundDefense
• Expected (ReDMCSB): F0321 sums F0313(slot) over the wound mask bits, averages, and scales attack by (130 - avgDefense) / 64. F0313 checks each worn item’s defense value for the specific wound slot.
• Actual (Firestaff): M10 and M11 build source-backed per-wound snapshots, use F0313 defence, then apply F0321’s (130 - avgDefense) / 64 scale.
• Evidence: dm1_v1_orch_pending_damage_flow and combat integration coverage.

BUG-102 FIXED — Fire/Spell Shield Defense Applied

• Severity: Resolved runtime
• Category: Mechanics
• Description: The former gap in F0321’s fire and magic branches is closed.
• ReDMCSB Reference: CHAMPION.C F0321:1842-1857 P0663_i_Attack -= G0407_s_Party.FireShieldDefense (C1_FIRE case) and P0663_i_Attack -= G0407_s_Party.SpellShieldDefense (C5_MAGIC case)
• Expected (ReDMCSB): After F0307 statistic adjustment, subtract party shield defense from attack value.
• Actual (Firestaff): combat_apply_defender_statistic_adjustment performs F0307 followed by the corresponding source party-shield subtraction; magic then takes the source no-armour-scale exit.

BUG-103 FIXED — Luck System Is Runtime-Bound in Combat

• Severity: Major
• Category: Mechanics
• Description: F0308’s two-stage random path and bounded Luck writeback are implemented in the combat resolver.
• ReDMCSB Reference: CHAMPION.C F0308:1120-1145
• Expected (ReDMCSB): Champions with high Luck stat have a chance to avoid hits. Cursed items with negative Luck exploit (BUG0_38 in original). Luck decrements by 2 on lucky roll, increments by 2 on unlucky roll.
• Actual (Firestaff): The live lifecycle Luck field is copied to the combat snapshot and written back after resolution. The PC/I34E non-positive-Luck branch is modelled; the unsafe negative-Luck exploit is intentionally not claimed.

BUG-104 FIXED — Creature AI Profiles Are Full-Tier

• Severity: Resolved runtime
• Category: Mechanics
• Description: The historical all-STUB state no longer reflects the runtime.
• ReDMCSB Reference: GROUP.C F0028/F0029/F0194/F0196/F0199/F0200/F0201/F0202/F0203, DEFS.H G0243_as_Graphic559_CreatureInfo
• Expected (ReDMCSB): Each creature type has specific movement patterns, attack behaviors, and special abilities (levitation, non-material, side attack, back-row preference, etc.).
• Actual (Firestaff): All 27 G0243 profiles are CREATURE_IMPL_TIER_FULL; the F0804 per-type dispatch owns poison, theft, ranged, levitation, non-material and archenemy branches. The STUB fallback remains defensive only and has no registered profile.
• Evidence: dm1_v1_creature_ai_behavior_source_lock.

BUG-105 FIXED — Creature Attack Ordering Is Source-Locked

• Severity: Resolved runtime
• Category: Mechanics
• Description: The old F0229 review marker was superseded by the source-owned ordered-cell table and dispatch bridge.
• ReDMCSB Reference: GROUP.C F0229:685-710
• Expected (ReDMCSB): Creatures attack party members in a specific cell-based priority order determined by the creature’s position and direction.
• Actual (Firestaff): F0229_DM1_GROUP_SetOrderedCellsToAttack_Compat resolves the original G0023 row and M10/M11 consume it for target selection.
• Evidence: dm1_v1_ordered_cells_to_attack_pc34_compat (222 assertions) and pass803_dm1_v1_ordered_cells_to_attack.

BUG-106 FIXED — Creature Flee Behavior Source-Locked

F0820_DM1_GROUP_GetFleeDirection_Compat() in src/dm1/dm1_v1_creature_ai_behavior_pc34_compat.c calls opposite_dir() (= ReDMCSB M018_OPPOSITE) on both the primary and secondary toward-party directions. This is the source-locked F0209 T0209094_FleeFromTarget formula. The F0201_GROUP_GetSmelledPartyPrimaryDirectionOrdinal helper (m10) is still a no-op as documented, but the caller-side opposite_dir() is correct. fearCounter in the FLEE state (memory_creature_ai_pc34_compat.c:1432) also decrements per the source’s T0209094 path.

BUG-107 FIXED — Thieves Eye Duration Is Source-Locked

• Severity: Resolved runtime
• Category: Mechanics
• Description: The earlier approximation was replaced by the F0412 source expression.
• ReDMCSB Reference: MENU.C:1960-1963 (T0412032 tail)
• Expected (ReDMCSB): T0412032 shifts the source spell power once and squares it before scheduling C73.
• Actual (Firestaff): F0757_MAGIC_ProduceOtherEffect_Compat uses spellPower >>= 1; durationTicks = spellPower * spellPower; and schedules the C73 timeout.
• Evidence: test_dm1_v1_mnu03_f0757_spell_duration_source_lock_pc34_compat plus the M10 status-timeline regression.

BUG-108 FIXED — Light Amount Table Source-Locked

The full 16-entry dm1_light_power_to_amount[16] = { 0, 5, 12, 24, 33, 40, 46, 51, 59, 68, 76, 82, 89, 94, 97, 100 } table is now source-locked in src/dm1/dm1_v1_light_pc34_compat.c (ReDMCSB DATA.C:359/1088). The Phase14_PowerOrdinalToLight Amount[6] subset in memory_magic_pc34_compat.c consumes the canonical table (indices 1..6) and matches exactly. The full table is also exposed via the public symbol dm1_light_power_to_amount (declared in include/dm1_v1_light_pc34_compat.h).

BUG-109 FIXED — Champion Stat Gain Cycle Source-Locked

F0331 is implemented in src/dm1/dm1_v1_champion_needs_pc34_compat.c (the post-M10 v1 split of CHAMPION.C). The per-tick stat recovery loop follows ReDMCSB CHAMPION.C F0331:2487-2497 (MEDIA240 branch = DM1 Atari ST 1.2+):

• BoundedValue(1, (MaxStamina » 8) - 1, 6) cycle count
• gain = (MaxHealth » 7) + 1 per cycle
• (current_stamina < max_stamina) consumes the gain
• time_criteria = ((GT & 0x80) + ((GT & 0x100) » 2) + ((GT & 0x40) « 2)) » 2
• Mana regen gated by wisdom + (wiz_skill + priest_skill)
• StaminaGainCycleCount starts at 4 (AL9995)
• Party-resting flag doubles all gains

• Negative food/water drain when below 0 is preserved

• Severity: Resolved runtime
• Category: Mechanics
• ReDMCSB Reference: CHAMPION.C:F0331 2362-2415 and 2487-2497
• Actual (Firestaff): dm1_v1_champion_needs_pc34_compat.c follows the source stamina-cycle loop, food/water thresholds, 64/16 outer cadence, rest multiplier, and the current -= current / maximum over-maximum statistic recovery rule. The former review-marker text was stale.
• Evidence: the champion-needs and clock-tick source-locked regressions; this remains distinct from broad campaign-capture parity.

BUG-110 NOT APPLICABLE TO DM1 — CSB Magic-Map State Is Isolated

• Category: Scope correction
• ReDMCSB Reference: DEFS.H:695-703, MENU.C:1873-1915, and TIMELINE.C:2002-2012 explicitly mark C80..C83 magic-map fields/events as Chaos Strikes Back-only.
• Actual (Firestaff): DM1’s 25-entry spell table admits no magic-map rows; CSB owns the per-champion counters and C80..C83 expiry handling in the lifecycle path. This is not a missing DM1 mechanic.
• Remaining work: CSB needs real-media magic-map capture coverage before this can earn broad CSB UI/runtime parity credit.

BUG-111 FIXED — Projectile Sub-Cell Hit Mask Source-Locked (partial)

m11_game_view.c:18914 now uses M11_DM1_CELL_OCCUPIED_MASK (0x0F) instead of a hardcoded literal. The mask is defined in include/m11_game_view.h as a ReDMCSB source-locked constant per DEFS.H M550 (DUNGEON.C:1085). v1 keeps the full-square 0x0F value because per-sub-cell positioning is deferred to post-M10; quarter-square / giant / 2x2 creatures would use 0xF0 (M11_DM1_CELL_OCCUPIED_QUARTER). The hardcoded literal was replaced with the named constant.

BUG-112 NOT AN ORIGINAL-SAVE FORMAT GAP — Field Masks Are Internal

• Category: Scope correction
• ReDMCSB Reference: LOADSAVE.C:F0433 writes source game/champion data as opaque byte blocks; F0435 restores those blocks. It has no per-field mutation-mask format.
• Actual (Firestaff): DungeonMutation.fieldMask is documented as a bounded internal replay abstraction. The original PC34 F0433/F0435 envelope is independently authenticated and round-tripped by the original-save handoff path.
• Remaining work: external original saves are still needed to prove broad interop across real campaigns; the supplied game archives do not contain a real save corpus.

BUG-113 FIXED — Poison Applies Vitality Adjustment

• Severity: Resolved runtime
• Category: Mechanics
• Description: Creature and projectile poison route through F0307 Vitality before mutation.
• ReDMCSB Reference: CHAMPION.C F0307 vs Vitality statistic
• Expected (ReDMCSB): Poison attack value is adjusted by defender’s Vitality via F0307_CHAMPION_GetStatisticAdjustedAttack before applying.
• Actual (Firestaff): The source-scaled poison value is used before the poison event is scheduled.

BUG-114 FIXED — Psychic Defense Uses Source F0030 Truncation

• Severity: Resolved runtime
• Category: Mechanics
• ReDMCSB Reference: CHAMPION.C:F0321 C6 attack branch and BASE.C:F0030.
• Expected: psychic attack uses ((attack * (115 - Wisdom)) >> 6) with the original integer truncation.
• Actual (Firestaff): F0762_MAGIC_GetDefenderPsychicAdjustedAttack_Compat now uses the same truncating scaled product. The regression boundary 5 * 100 >> 6 == 7 rejects the former half-up result of 8.
• Scope: DM1 does not expose a player psychic spell, but Ghost/Rive and Screamer creature attacks consume this F0321 C6 defense route.

BUG-115 FIXED — F0306 Stamina-Adjusted Value Compiler Order Hazard

• Severity: Resolved runtime / source-package parity
• Category: Mechanics
• Description: The BUGX_XX compiler-order hazard at CHAMPION.C:1095 makes (P0641_i_Value >>= 1) + (P0641_i_Value * L0925/L0926) package-dependent: it depends on the original compiler, not on dungeon bytes or a user option.
• ReDMCSB Reference: CHAMPION.C F0306:1078-1103, BUGX_XX comment
• Expected: Megamax Atari ST and High C FM Towns evaluate the first operand first; Turbo C++ 1.01 DM1 PC 3.4 and Aztec-built DM1 Amiga 2.x evaluate the second operand first.
• Actual (Firestaff): GameWorld_Compat.pc34F0306FirstOperandFirst is bound at native DM1 handoff from verified Atari ST or FM Towns media. The M10 combat route and M11 shield-defense route call the source-selected helper directly. PC 3.4 and Amiga 2.x retain second-operand semantics.
• Evidence: dm1_v1_f0306_stamina_pc34_compat pins both results without the process-global test toggle; real Atari archive boot remains covered by dm1_v1_atari_st_outer_archive_real_data.
• Impact: Corrects low-stamina strength, load and shield-defense values on affected original packages.

BUG-116 FIXED — Runtime Dynamics Table Source-Locked

memory_runtime_dynamics_pc34_compat.c no longer carries a NEEDS DISASSEMBLY REVIEW marker. The active-group cap follows ReDMCSB GROUP.C:512-520 strictly (party-map only, no adjacency suppression). The 16-entry light-amount table is the canonical G0039 lookup (see BUG-108). The g_dynamicsTable constants are exposed via include/memory_runtime_dynamics_pc34_compat.h for FIRESTAFF_DATA tests to assert against.

BUG-117 — Test Infrastructure: Python Verification Scripts Hardcode Build Path

• Severity: Minor
• Category: Testing
• Description: At least 14 Python verification scripts in tools/verify_pass*.py search for test binaries in ROOT / "build" and ROOT / "build-*" but do not search the CMake build directory when it’s on a different volume (e.g., /Volumes/Extern-disk/legacy-workspace-work/firestaff/build). This causes 14+ test failures that are not real code bugs.
• ReDMCSB Reference: N/A (test infrastructure)
• Expected: Tests should find binaries in the actual CMake build directory.
• Actual: Tests fail with FileNotFoundError or AssertionError: missing built test_*.
• Impact: False test failures mask real issues; CI may pass because it builds in-tree.
• Fix Complexity: Low — add the external build path to the search candidates, or respect a FIRESTAFF_BUILD_DIR environment variable.

BUG-118 FIXED — Viewport Occlusion Gate Chain Root Failure (bounded)

pass434_dm1_v1_original_viewport_crop_readiness_gate is no longer the root failure: the bounded F0128 helper m11_dm1_v1_f0128_compose_viewport_for_tuple drives the readiness flag after every party-tuple change. The viewport-crop ready state (g_f0128_ready) is exposed via m11_dm1_v1_f0128_viewport_crop_ready(). G0076_B_UseFlippedWallAndFootprintsBitmaps toggle (m11_dm1_v1_f0128_g0076_set/get) implements the ceiling/floor alternation flag.

Source-locked per ReDMCSB DUNVIEW.C F0128 (8318-8611) + F0674F0128_sub (2995-2996). v1 keeps the bounded form: the M11 caller drives the helper after every party-tuple change; the actual bitmap copy is delegated to the existing M11 wall path (m11_dm1_v1_dungeon_compose g0296) which already calls F0674 via the existing wall path. The bounded readiness signal closes the gate chain root failure.

Regression gate: tests/test_dm1_v1_f0128_viewport_pc34_compat.c (7/7 PASS).

BUG-119 FIXED — Same-Depth Side Content Survives Center-Block Replay

• Severity: Major visual parity
• Category: Viewport / F0128 ordering
• ReDMCSB Reference: DUNVIEW.C:F0128 dispatches each depth in the order DnL, DnR, then DnC (D3 at 8491-8502, D2 at 8514-8525, D1 at 8527-8542).
• Former Firestaff behavior: the deferred M11 side-content pass rejected side cells at the same depth as the nearest closed center wall or door. That contradicted the source ordering and could make a side floor ornament disappear when approaching its depth.
• Actual: only side content farther than the nearest center blocker is rejected. Same-depth side cells remain eligible; the subsequent source-material structural replay provides the real overlap/occlusion.
• Evidence: test_dm1_v1_viewport_3d_pc34_compat source-lock assertion.

Summary Statistics

Priority Fix Order

1. BUG-112 (Minor) — preserve unproven save-field masks as a fail-closed limitation
2. BUG-110 (Minor) — verify per-champion magic-map state against a real save/trace

Fixes Applied in v2.7.13

BUG-101 FIXED — Armor Defense Now Uses F0321 Wound Defense Calculation

Replaced the skill-level approximation with proper ReDMCSB F0321 wound defense:

• Iterates all 6 champion wound slots (hands, head, torso, legs, feet)
• Looks up actual equipped armor defense values from things data
• Applies vitality contribution per wound slot
• Uses G0050 wound defense factor scaling per slot
• Applies the (130 - avgDefense) / 64 scaling formula

BUG-102 FIXED — Fire/Spell Shield Defense Now Applied

Added proper attack-type branching per F0321:

• FIRE attacks: subtract FireShieldDefense, then apply armor scale
• MAGIC attacks: subtract SpellShieldDefense, skip armor scale (goto T0321024)
• PSYCHIC attacks: skip armor scale
• NORMAL attacks: skip entire F0321 body
• BLUNT/SHARP/LIGHTNING: fall through to armor scale only

BUG-113 FIXED — Creature Poison Now Applied with Vitality Adjustment

Added creature melee poison application:

• Checks creature profile’s poisonAttack value
• 50% chance of poison per hit (matching ReDMCSB F0230:1395-1404)
• Poison value adjusted by defender’s Vitality via F0307 formula
• factor = 170 - vitality; if < 16: poison >> 3; else: (poison * factor) >> 7

BUG-117 FIXED — Test Build Path Fully Portable

Added tools/firestaff_build_dir.py shared helper with find_build_dir() / resolve_build_dir(). Lookup order:

BUG-101 FOLLOW-UP FIXED — Pending Damage Flow Now CTest-Gated

Added data-free CTest dm1_v1_orch_pending_damage_flow around F0889_ORCH_ApplyPendingDamage_Compat. The gate verifies the M10 handoff from GameWorld_Compat.pendingCombat into champion HP, wound bits, pending-combat clear, EMIT_CHAMPION_DOWN, and partyDead. This proves the end-to-end data-layer damage flow without relying on M11 rendering, real assets, or original-DOS capture.

1. FIRESTAFF_BUILD_DIR env var (already supported by some scripts; now standardised).
2. <root>/build (in-tree single-config).
3. <root>/builds/<cfg> (in-tree multi-config, iterates).
4. <root> parent walk for any CTestTestfile.cmake.

Updated 211 tools/verify_pass*.py scripts to import resolve_build_dir and replace Path(os.environ.get('FIRESTAFF_BUILD_DIR', str(ROOT/'build'))) with resolve_build_dir(ROOT, ROOT/'build'). 148 scripts also needed import sys added.

tools/test_firestaff_build_dir.py 6/6 PASS (env var, in-tree, multi-config, no-build fallback, resolve-fallback, real Firestaff project).

Commit 887ed7cb3.

BUG-119 FIXED — Champions die in Hall of Champions

Reported by user testing v2.7.13. When the C040 mirror candidate panel is open, the party can attack and kill the candidate creature before recruiting it. The candidate is rendered as a portrait graphic in the D1C cell, but the F0735_COMBAT_ResolveChampionMelee_Compat path had no invulnerability check.

Fix: added isCandidateInvulnerable flag to CombatantCreatureSnapshot_Compat and a guard in F0735 that bounces to COMBAT_OUTCOME_NO_ACTION when the flag is set. ReDMCSB CLIKCHAM.C F0367 lines 24-25: the candidate panel owns the front cell while selection is pending. New regression test: tests/test_dm1_v1_hall_of_champions_pc34_compat.c.

BUG-120 FIXED — Slow after selection in Hall of Champions

Reported by user testing v2.7.13. After picking a champion in the C040 panel, the game ran noticeably slower. The C040 panel was re-rendering the wall-ornament graphic every frame via m11_draw_dm1_front_mirror_route even though the panel chrome was already showing.

Fix: in m11_draw_dm1_front_mirror_route (m11_game_view.c:13479), added an early-return when state->candidateMirrorPanelActive is set. The wall- ornament blit is skipped while the panel owns the front cell; only the candidate portrait is drawn. This also fixes BUG-121 (the peach placeholder box is hidden during selection).

BUG-121 FIXED — Graphical artifacts in Hall of Champions

Reported by user testing v2.7.13. The C040 candidate panel showed an orange/peach placeholder box (wall-ornament graphic) floating in the D1C front cell next to the candidate portrait.

Fix: same as BUG-120 — the candidateMirrorPanelActive early-return in m11_draw_dm1_front_mirror_route skips the ornament blit while the panel is open, leaving only the candidate portrait graphic visible.

BUG-122 FIXED — Floor-object moves could lose exclusive ownership

• ReDMCSB reference: DUNGEON.C:F0163 (line 1769) links a THING at the end of one list, while F0164 (line 1840) removes its current list owner. MOVESENS.C:804→870 demonstrates unlink-before-link in a source movement path.
• Former Firestaff behaviour: the compact floor-cell representation could append an object to a destination without removing its old cell entry. A full destination could also update coordinates without granting a new list owner. That could manifest as duplicated, disappearing, or apparently thrown items during inventory/floor interaction.
• Fix: DM1_V1_Object_DropPc34Compat now performs unlink → capacity check → link, restoring the old list member exactly if the destination is full. Pickup rejects a coordinate whose cell does not own the object, rather than manufacturing a carried object.
• Evidence: dm1_v1_object_interaction_source_lock; authentic Atari ST 1.2 and Amiga 2.0 handoff tests cover the original-media alcove object.

Notes

• BUG-110 (Magic Map) and BUG-112 (Savegame fields) remain explicit validation limits. BUG-103, BUG-104, BUG-105, BUG-106, BUG-107, BUG-109 and BUG-111 are covered by current source-locked runtime routes; they are not backlog items.
• The two recent fixes in v2.7.12 (wall texture behind mirrors f99587c35, FTL swoosh palette e39c3d804) are verified correct and not re-opened.
• Phase A probe passes 23/23 invariants.
• The M10 data layer (memory_combat_pc34_compat.c) has a well-structured F0321 implementation that could replace the inline approximation in m11_game_view.c.