Session handoff 2026-08-02: Vulcanus cliffs at 0.9720/0.9713 - next lead is the 37 one-edge orientation errors

[wormeyman]

I'm resuming work on FactorioMapWebUI in /Users/ericjohnson/GitHub/FactorioMapWebUI (branch main, clean tree, HEAD e68f617).

Context

Static Vue 3 SPA for authoring Factorio map-generation presets, plus a client-side preview that ports the game's noise/autoplace generation to TypeScript. Current goal: get Vulcanus cliff accuracy to 100%, tracked in issue #84.

Issue #84 stays OPEN until Vulcanus cliffs are 100% or Eric says to close it. Do not close it for having an empty item list - re-scope it instead. (Eric's instruction, 2026-08-01. It was auto-closed once already: a commit message reading Closes #84 item 4 closed the whole issue, because GitHub parses Closes #84 and drops the qualifier. Write "closes item N of #84" instead, and check gh issue view 84 --json state after any merge.)

Current accuracy

	recall	precision	wrong orientation
Nauvis	1.0000	1.0000	0 / 334
Vulcanus (shipping)	0.9720	0.9713	37 / 1531 = 2.4%

Per region: [0,0] 283/283, [1500,1500] 885/900, [-1200,800] 401/387. Session start was recall 0.806/0.938/0.853 with 12.5% wrong orientations.

What's done this session (all merged to main)

• e027566 (test(cliffs): the "excess" was the lava rejection, not a second defect (#84) #86) - Vulcanus cliffs: the 2.4% remaining after the multisample grid fix #84 items 1 and 2 were one measurement error: our side never ran tryToAddCliff's lava-collision rejection while the game always does. 185 of the "187-cell excess" was that unapplied rule. Precision 0.8719 -> 0.9743.
• b843eaf (test(noise): channel audit - cliffs ARE the only coarse-grid consumer (#84) #87) - channel audit (Vulcanus cliffs: the 2.4% remaining after the multisample grid fix #84 item 4) closed. multisample appears in exactly one place in all of factorio-data, so the audit reduced to "who consumes Vulcanus elevation": cliffs (grid 4), tiles+temperature (grid 1, measured), and nothing else - resources and rocks contain 0 references to elevation.
• e68f617 (test(cliffs): pin the SHAPE of the orientation residual (#84) #92) - pins the SHAPE of the orientation residual as a permanent test.
• 8243b65 (fix(cliffs): rotbb collision boxes are ROTATED, not their bounding box (#84) #88) - WRONG, superseded by fix(cliffs): the collision box is the RAW rectangle - #88 was wrong (#84) #90. Claimed the collision box is rotbb's rectangle rotated 45 degrees.
• 64b5b79 (fix(cliffs): the collision box is the RAW rectangle - #88 was wrong (#84) #90) - the correction. Disassembly shows the engine uses the RAW stored rectangle and discards the box's 1/8 orientation tag.

The most important lesson of the session

#88's model scored best on every metric and was wrong.

box	false rejections	recall	precision	evidence
AABB (until #88)	13	0.9675	0.9743	an assumption
45-degree oriented rect (#88)	0	0.9758	0.9727	empirical fit only
raw stored rect (#90, current)	6	0.9720	0.9713	disassembly + API docs

#88 hit 13/13 by shrinking the box past what the engine uses, and the excess shrinkage also absorbed an unrelated defect: 4 of the 6 cliffs the correct box still rejects are cells whose orientation we get wrong, so we load the wrong box entirely. A model that scores perfectly by hiding a second bug is worse than one that leaves it exposed. When a fix lands on a metric perfectly, treat that as a prompt to find independent evidence, not as the evidence. The same failure happened the day before in #86.

Next task

Two leads remain, both well characterised. Take the orientation one first.

1. The orientation residual - 37 cells, ONE edge each, one-directional

Measured tonight:

• All 37 differ in exactly one edge. Not one two-edge difference. (Before fix(cliffs): multisample offsets are in GRID UNITS - #18 root cause #83 it was 125 of 175 at two edges.)
• Every one runs the same way: the game says -to-none, we say a crossing. So we over-detect a crossing on one edge, 37 times in ~6100 edge evaluations (0.6%), and never under-detect.
• Per region 7 / 26 / 4. Which edge: L11 / R6 / T7 / B13 - no directional bias.

Two causes are already ruled out - do not re-derive them.

crossesCliff is exact. I disassembled CliffGenerator::crossesCliff (0x10160c914 under 2.1.12 - the notes' 0x101606d08 has moved) and src/noise/cliffs/cliffPlacement.ts reproduces it line for line: the a < 0 || b < 0 early-out, the band derivation, the boundary < e0 check, the strict cliffAvg > 0.5 gate and the strict crossing comparisons. There is no >= vs > slip to find.

cliffiness_basic is exonerated. Substituting the game's own corner cliffiness (from oracle-vulcanus-cliff-corner-fields-entity-regions.seed123456.json) leaves the count at exactly 37 / 1531 - no change at all. Control: substituting the game's 1-tile elevation from the same fixture gives exactly 175 wrong / 1400 matched, the pre-#83 figures to the digit, which proves the harness is live and re-confirms #83.

So the defect is in the grid-4 cliff-elevation field - the one input in the chain with no direct per-corner oracle. #83 established that channel with a contour-shift experiment, never a per-corner comparison.

The shape is pinned in test/cliffOrientationResidual.spec.ts (e68f617): one edge, always over-detecting, spread across edges and regions. If under-detections ever appear there, the cause has changed.

Concrete next step: capture the game's cliff_elevation at the 4-tile corner lattice around the 37 disputed cells, using sampleCliffEntitiesFull's probeExpression + probeProperty (register an expression at the data stage and route it onto cliff_elevation, so the cliff generator becomes the readout rather than calculate_tile_properties). Compare per corner. A small offset there would explain a one-directional over-detection.

2. Over-placement below elevation 120

test/vulcanusElevationLevels.spec.ts sweeps cliff_elevation_0; the low/high ratio gap is 0.024 (worst-low 1.0415 at level 40, worst-high 1.0177 at level 130). Pure over-placement now, no recall cost. Note this figure has been reported three times today (0.067 -> 0.018 -> 0.024) as upstream fixes landed - treat it as provisional.

3. Perf (#84 item 3) - untouched

pnpm perf is FMW_PERF=1 vp test test/render-cost.perf.spec.ts. It must run on a quiet machine - it is wall-clock render timing, and this repo has already recorded a 3001ms baseline reading 16729ms on identical code under load. Do not run it alongside other work, and do not run it in parallel in a worktree; a worktree isolates files, not CPU.

Read these first (authoritative - don't re-derive)

• /Users/ericjohnson/GitHub/FactorioMapWebUI/CLAUDE.md - project conventions. Read fully before touching anything.
• /Users/ericjohnson/GitHub/FactorioMapWebUI/docs/noise/vulcanus-cliffs-NOTES.md - the STATUS banner has current numbers. The last three sections each correct the one before it - read all three in order or you will act on a superseded number. The final one, ## The collision box, settled by disassembly, is current.
• /Users/ericjohnson/GitHub/FactorioMapWebUI/docs/noise/cliffs-NOTES.md - STATUS banner, then ## ROOT CAUSE, 2026-08-01.
• gh issue view 84 - the itemised remainder plus tonight's two analysis comments.
• Claude Code memory at /Users/ericjohnson/.claude/projects/-Users-ericjohnson-GitHub-FactorioMapWebUI/memory/ - MEMORY.md is the index. Most relevant: cliff-collision-box-is-raw-rect.md, comparison-must-run-the-same-rules.md, keep-84-open-until-100.md.

Conventions / gotchas

• pnpm run verify is the gate (vp check + check:vue + vp test + preview:test), ~65-90s local, ~4 min CI. pnpm vp check --fix for format+lint+types. Never add a tsc typecheck script.
• Branch for work; PR and merge with gh pr merge <n> --squash --delete-branch=false. Do not commit to main directly.
• pnpm refs:sync --check before trusting any reading of game data. Currently in sync at 2.1.12.
• Disassembly is cheap and beats tuning models against a metric - the binary is unstripped. Recipe: nm -n "$BIN" | c++filt for the VA, then lldb -b -o "disassemble --start-address <VA> --end-address <VA>" "$BIN". The --name form silently returns nothing on many symbols. VAs move between builds - always re-derive from nm. Binary at ~/Library/Application Support/Steam/steamapps/common/Factorio/factorio.app/Contents/MacOS/factorio.
• Also grep factorioLuaAPI/ (API docs) and ~/GitHub/factorio-data (prototype Lua) before WebFetching anything Factorio.
• Adding a fixture means adding its test/fixtures/PROVENANCE.json entry - a test enforces it.
• git checkout <file> is not an undo for uncommitted edits. I wiped a fix that way tonight while reverting a planted failure. Copy to the scratchpad or commit first.
• Assert non-vacuity on every measurement. Two silent-zero results tonight came from a hand-derived helper whose fields were undefined; both looked like clean findings.
• Temp/scratch files go in the session scratchpad, never /tmp or the repo.
• Put decisions to Eric via an interactive prompt with grounded options, not prose questions.

Also open (not this thread)

#24 cliff/ore separation, #22 approximations register, #16 cross-overlay occupancy, #27 Fulgora island finder, #58 Renovate dashboard.

[wormeyman]

Update before shutdown: two further results landed after this handoff was written, both narrowing the orientation lead in #84.

1. crossesCliff is exact. Disassembled at 0x10160c914 (the notes' VA had moved) and cliffPlacement.ts reproduces it line for line — negative-corner early-out, band derivation, boundary < e0 check, strict > 0.5 gate, strict crossing comparisons. The >= vs > hypothesis in the 'Next task' section above is refuted; don't spend time on it.
2. cliffiness_basic is exonerated. Feeding the game's own corner cliffiness changes nothing — same 37 wrong, same 1531 matched.

So the residual is isolated to the grid-4 cliff-elevation field, the one input in the chain with no direct oracle. The next step is the capture described in the handoff (probeExpression + probeProperty onto cliff_elevation at the 4-tile corner lattice), and it is now the only remaining candidate rather than one of three.

Also worth carrying forward: substituting the game's 1-tile elevation reproduces exactly 175 wrong / 1400 matched — the pre-#83 figures to the digit — which independently re-confirms #83 and proves the substitution harness is live.

[wormeyman]

Handoff picked up and its stated lead refuted, in PR #93 (merged, fb7d899).

The next-task section here said the residual was isolated to the grid-4 cliff-elevation field and the step was to capture cliff_elevation at the 4-tile corner lattice. That field is now measured through the cliff generator itself and it is exact - as is cliffiness_basic (over its unclamped interior), the smoothing stencil on both axes, cliff_smoothing = 1 (read back off the game rather than inferred from the prototype default), crossesCliff, crossingsForChunk and fixImpossibleCells.

The residual is two defects, not one. With cliff_smoothing = 0 and everything else real: [0,0] and [-1200,800] are exact, [1500,1500] is still 21-wrong. Scoring only [0,0] - which is what this handoff's framing implied - concludes "it's the smoothing", and that is false for 21 of the 37.

Also re-scoped: #84 item 2's "over-placement below elevation 120" is one contiguous blob where the game places no cliff under any field routed onto it. That is now the sharpest lead.

Full detail in #84's comment and docs/noise/vulcanus-cliffs-NOTES.md (## The residual is TWO defects, and the field is not either of them). #84 stays open.

Closing this handoff.