Skip to content

BACKLOG.md

Latest commit

 

History

History
203 lines (183 loc) · 12.2 KB

BACKLOG.md

File metadata and controls

203 lines (183 loc) · 12.2 KB
Raw

Backlog — parser / Elite Insights parity

Historique — pas une source de priorité actuelle. Ce score et ces pistes datent du 2026-08-07. Conserver comme provenance de parité, mais repartir du contrat accepté, du checkpoint courant et d'une mesure reproductible avant de sélectionner un travail.

Measured with uv run python scripts/ei-parity/ei_diff.py over the 35-log corpus. Setup and probes: docs/ei-parity-workbench.md. Findings so far: docs/parser-audit-2026-07-31.md.

Scoreboard: 444 differences over the 35-log corpus (2026-08-07). Per-session history in SESSION.md.

Trajectory: 23 830 (2026-07-31 audit) → 4 503 → 1 116 → 798 → 711 → 690 → 644 → 635 → 632 → 583 → 533 → 489 → 444.

rotation is reported twice: a bucket count of player rows, and the cast counts under the total. Judge any rotation change on the casts — a partial finder can degrade the data while leaving the row count untouched.

Note the "0 diffs corpus-wide" figure quoted in #117/#118 was measured over 20 logs. The committed corpus is 35; always run the harness with no arguments.

Done (2026-07-31)

  • Reproducible local EI ground truth (EI 3.26.0.0 CLI + .NET 8, offline)
  • Stratified 35-log corpus spanning ten arcdps builds
  • Correct elite-spec catalogue (EliteSpec, _VALID_ELITE_BY_PROFESSION)
  • Anonymized enemy names compared in English, not the client's language
  • Drop the fabricated result = 13 on zero-magnitude condition ticks
  • Decode the condition result enum per arcdps era (2026-05-07 rebase)
  • Read overstack_value (barrier absorption / EI shieldDamage)
  • Fold the 2026-05-07 buff statechanges (69 / 71 / 72) back onto the apply/remove path — boon uptimes were ~0 on every recent log
  • Fold the 2026-05-07 cast statechanges (67 / 68) back onto the activation path — rotations were missing every non-instant cast
  • Resolve an EI target to every agent sharing its instance ID
  • Split condi/power by EI's buff classification, not the arcdps channel

Done (2026-08-04)

  • Root cause of the per-target over-attribution (#120). EI emits one player entry per contiguous stretch of squad membership, not one per player: an account appears several times with adjacent firstAware / lastAware windows and a different group on each, and every counter on an entry covers only that stretch. Each entry is now compared against its own slice of the event stream. 4 503 → 1 140.
  • Consumables are no longer sliced (#120 follow-up) — they are pre-fight applications EI repeats on every entry. 1 140 → 1 116.
  • postcss raised past GHSA-fxqj-rqcc-2cmp (#119). The overrides pin in web/pnpm-workspace.yaml was itself what blocked Dependabot.
  • Buff simulation stops at an actor's last-aware (#130). 798 → 779.
  • Down-contribution row sliced per EI entry (#132). 779 → 711.
  • teamID reported only on an account's last entry (#134). 711 → 690.
  • Backend→frontend drift check (#121). PlayerProfile was missing boon_strips / condition_cleanses; PerFightBreakdownRow was missing 28 more. scripts/ei-parity/api_coverage.py now guards both directions.

Done (2026-08-06)

  • Four instant-cast finders transcribed from EI's sources, each verified exact over the corpus before being wired: 9084 "Advance!" (self-aegis 20–40 s), 5535 Cleansing Fire (by-dst + two secondary effects), and the four Evoker familiar skills (credited to the owner). 266 casts recovered, 0 introduced. 690 → 644. scripts/ei-parity/probe_ei_finders.py scores a candidate rule.
  • SkillActivationEvent carries src_master_instid, so a familiar's cast can be credited to its owner without re-attributing the cast.
  • Three missing engineer kits (Bomb 5812, Tool 5904, Grenade 6020). EI declares seven EngineerKitFinders and reads each bundle off /v2/skills; we had four. 644 → 641.
  • 30961 Exit Reaper's Shroud now fires only on a full removal and lands at min(swap - 1, time), matching BuffLossCastFinder + UsingBeforeWeaponSwap. 45 fixed on each side. 641 → 635.
  • UsingNoAnimatedCastChecker ported properly for the guardian symbol traits (13684, 13677) — a cast-window test, not "a cast is open right now". 19 spurious casts dropped.
  • BuffStackActiveEvent now reaches the buff tracker. The Activate path existed but ei_compare never delivered the events, so regeneration's queue was never reordered. 185 → 141 buff diffs.
  • The two Activate overloads separated. Only the explicit stack-active record gets EI's richer rule (replace a nearly-spent active stack, pin noSort); an addedActive apply just moves its stack to the front. noSort is now tracker-wide, matching EI's shared static HealingLogic.
  • Regeneration evicts the stack arcdps named (scan_regeneration_overstacks + HealingLogic.FindLowestValue). The uncredited removals that record which stack an application displaced are dropped from the event stream — as they are in EI — but recovered as a side input. Regeneration 116 → 17 over the session. 583 → 533.
  • againstDownedCount asks whether the hit landed, not whether it hurt. A hit absorbed or converted to barrier lands for zero and still counts. All 45 differences resolved; the shared predicate now lives in damage_predicates.landed_hit. 489 → 444.
  • A down that coincides with the death earns no contribution. EI keeps a downed segment only when start < end, so such a down has no segment and IsDownBeforeNext90 returns false for every hit that led to it. 533 → 489.
  • Base attunement skills are no longer booked for a Weaver. EI books weaver dual-attunement swaps as separate skills and never the base one; confirmed corpus-wide on all four elements. 36 spurious casts dropped.

Ruled out — do not retry

  • Windowing buff uptimes per slice. EI reports the same whole-fight buffUptimes on every entry of a split account; slicing them takes buff differences from 225 to 913 on the four logs that have splits.
  • Subtracting two BuffStateTracker readings to get stack-time inside a window. The tracker only moves forward, so both readings taken after the full feed return the same total. Replaying per window computes the window correctly and still diverges, because the target is whole-fight.
  • Shifting consumable timestamps by +1 ms. 39 mismatching entries are short by exactly 1, but most entries already agree, so the blanket shift takes consumables from 39 to 95.
  • Narrowing a target back to one agent per instance. Regresses 20260224 from 682 to 3 205.
  • Emitting 9084 "Advance!" for every guardian shout effect with fewer than five self-stabilities. Only 29 of 70 such effects are real casts; the other 41 become false positives. Corpus 690 → 726 (8 fixed, 44 introduced). The 50 ms instant-cast ICD does not absorb them. Superseded 2026-08-06: the real rule is a self-applied aegis of 20 to 40 seconds, and it is exact.
  • Deriving an InstantCastFinder by correlation. Three attempts failed the same way: a trigger that covers every cast is not one that fires only on casts, and a single log cannot tell them apart. Both triggers found this way were wrong on inspection of EI's sources — 77370 Zap is a minion finder, not BoonApplyEvent 76639. Read the rule from .tooling/ei-src, then verify it with probe_ei_finders.py.

Next (2026-08-06, in impact order)

  • players.rotation (308 rows / 668 casts missing, 215 extra) — still the largest bucket, and the method is now settled: transcribe the finder from .tooling/ei-src (a shallow sparse clone of baaron4/GW2-Elite-Insights-Parser, MIT), then verify it with scripts/ei-parity/probe_ei_finders.py before wiring anything. Adding a rule to that probe is a few lines. Current heads of the distribution: 29560 Spiteful Spirit (30) is the head of the missing side and is not explained by either of its two declared finders — see SESSION.md; it needs a path outside NecromancerHelper.
  • Weaver dual-attunement swaps. EI books sixteen of them as their own skills: DualFireAttunement (43470) and its three siblings carry real ids, the twelve mixed ones carry synthetic negative ids (-5 through -20, see SkillIDs.cs). It derives them from the pair of attunement buffs a weaver holds (WeaverHelper.GetLastAttunement). We now drop the base-attunement casts for weavers rather than mislabel them, so these sit squarely in the missing column.
  • buffUptimes (86) — might (740) is now the head at 23, then regeneration 17, swiftness 13, stability 9, fury 7, aegis 6. Regeneration is close to done: 11 of its 17 are under 0.1 and only 2 exceed 2 points. Might is an intensity buff — average stacks, not uptime — so it runs through the other branch of _advance and the OverrideLogic stacking rules rather than HealingLogic. Read EIData/Buffs/BuffSimulators/BuffSimulatorNoID/EffectStackingLogic/OverrideLogic.cs before touching it.
  • Read buff capacities from the log, not a table. EI overrides its hard-coded capacity with the BuffInfo record's MaxStacks (statechange 30, SrcMasterInstid). Our _CAPACITIES agrees with the corpus on every tracked boon except alacrity (we default to 9, the logs declare 99). No current difference depends on it, so this is correctness rather than parity.
  • group (7), name (3), and the single-digit dps / defenses / appliedCrowdControl residuals — the long tail is now 35 differences across a dozen buckets, each likely its own small rule.
  • downContribution residual (9 + 6) — the inst 7121 oddity recorded here turned out to be the down-on-death rule and is fixed. What is left has not been characterised.
  • group (7), consumables, float residuals — long tail.

Still open from earlier passes

  • Refresh the visual-regression baselines on Linux. #121 moved the player profile by 86 px, so 06-player-profile-with-timeline.png and 07-player-empty-timeline.png are stale. macOS Chromium does not reproduce the committed captures, so this must run on Linux: cd web && UPDATE_BASELINES=1 pnpm exec playwright test --project=visual-regression. Baselines live in docs/screenshots/.
  • Render the 28 per-fight boon columns. They reach the client (#121) but nothing displays them; 28 extra table columns needs a design.
  • Recorder-agent surplus on 20260125-194936 — for krill le faucheur.1679 skill 9107 we emit 28 records where EI reports 16 hits. Most of the surplus carries is_offcycle = 1 at distinct timestamps, so they are not literal duplicates. Reproduce with scripts/ei-parity/probe_raw_events.py 20260125-194936 "krill le faucheur.1679" 9107.
  • overstack_value on direct-damage records — a direct hit carries overstack = 2265124730, which is not a barrier amount. shield_damage only guards the INT32_MAX sentinel, so garbage just under the cap would leak. Nothing compares shieldDamage yet, so this is latent.
  • consumables — a mix of a +1 ms offset on some entries and entries we never emit. extract_initial_buffs is used only by ei_compare and its own test, so it can be changed freely.

Test-suite hygiene

  • test_parser_multilog_corpus.py and test_parser_dps_report_golden.py point at /home/roddy/... and skip everywhere else. Move to a small committed fixture so they actually run in CI.
  • Replace assert len(differences) <= N with exact-match assertions on the buckets already at zero. A ratcheting threshold cannot distinguish "fixed" from "not yet broken enough to notice".
  • apps/api webhook tests need a live Postgres and error out locally. Either mark them integration or provide a compose target.