- R1 soft delete: add deleted_at to slabs; exclude deleted rows from GETs
- R2 publish: terminal poll returns woo_product_id then the working row is
discarded; the app never deletes from WooCommerce (reworks PUBLISH-25)
- R3 prompts: SQLite is the point of record; file-based prompts superseded,
compose prompt mount removed
- R4 images: photo bytes are SQLite BLOBs; /images is a health probe + staging
Details recorded in .audit/m3-build-plan.md.
- PUBLISH: attempt-level failure wording, no image cap, failure triage rows
- METADATA: title rule is fallback-only (Call 2 not run)
- REVIEW: calculated widths/sqft/bdft not user mutable
- IDs now step by 5 per namespace; rule recorded in STRUCTURE.md
Reconciles the scattered requirements onto the owner-job namespaces.
- docs/REQUIREMENTS.md: 128 requirements across CAPTURE, METADATA, REVIEW,
PUBLISH, SETTINGS, APP, OPS, QUALITY
- docs/STRUCTURE.md: Test may be a path, a requirement ID, or - ;
a shipped path must exist
- tools/check_requirements.py: resolve test refs accordingly
- .audit/2026-09-18-reconciliation-decisions.md: full decision record plus
18 deferred rulings for the final cleanup pass
Gate: python tools/check_requirements.py -> PASS
Copy listing-copy/Woo payload spec from main. Lock {character} to the
primary Figure term name (Cathedral example). Lock categories[] to all
ready-set species and wood-category ids.
List the five locks from comparing CREATE.md to main technical docs
(categories payload, attributes shape, overlay upload, character
template source, server persistence).
Reorganize docs for humans: README map (goals / shipped / M3), PRODUCT
overview, and a specific CREATE.md that pulls Woo sync, metadata, pricing,
and inference requirements from main. Exit criteria is a slab submission-
ready for WooCommerce, proven with a SLAB-UAT-* draft. Add UAT-PRACTICE.
Add docs/CREATE.md as M3 SoT from Ty conversation (Woo attributes sync,
ready-set use, create, OpenAI Settings dependency). Point AGENTS/PRODUCT
here; mark main milestone splits obsolete for this branch.
Lower the axial body threshold while retaining fringe rejection, minimize length-only tip inset, and make sheet sensitivity curves testable and conservative at zero.
PhotoStage used new Image() while mask worker and calculate used bare
createImageBitmap; Orientation-8 matched-set JPEGs can disagree across
those paths so the approved pink axis is not the overlay (ux,uy).
One decodeOrientedBitmap helper always requests imageOrientation
from-image (with bare fallback), wired into PhotoStage, mask.worker,
and session calculate/siblings.
Root cause: 0.2%/99.8% axial percentiles still included dense fringe past
the visible body; round line caps and scale-tied font overshot edges and
clipped width labels. Use across-width tip stations, inset length/width
ends for stroke/arrowhead, butt caps, inventory-capped font, axis-right
length label, clamp labels to PNG.
Length dim already mask tip-to-tip (locked by test). Width labels move
axis-left off the wood via labelAt; font steps down to max(22, scale*1.0).
Inventory/overlay crop FILL 0.8→0.9; CAPTURE/PHOTOS docs match.
Lock length dim to mask tip-to-tip (not oversized bbox), require width
labelAt on axis-left off the wood, smaller font contract, and 90% crop
fill extent. Red before green.
Add buildPhotosReadyLineup (clean cal → overlay → batch inventory PNGs
in capture order, plus empty-mask/floor error slots). MeasurePanel
switches to a Photos ready end screen with Milestone 1 numbers and the
inventory PNG lineup; Back still returns to Length axis.
Failing TDD for M2 Photos-ready end screen: pure buildPhotosReadyLineup
orders clean calibration inventory PNG → dimension-overlay copy → other
batch inventory PNGs in capture order (skipping cal null), surfaces
batchApply empty-mask/floor errors, and MeasurePanel must show Photos
ready with Milestone 1 numbers (not source thumbs).
Add pure station/axis math and canvas burn-in for a second calibration
inventory PNG (results.overlayPng). Width stations start 6 in from the
EXIF-oriented photo bottom along the approved length axis (6, 18, … ≤
L−6); length graphic sits on axis-right outside the slab. Clean
inventoryPng and batchApply[cal]=null stay unchanged.
Failing TDD for M2 dimension-overlay copy of the calibration inventory PNG:
width stations every 12 in from EXIF bottom (6, 18, … ≤ L−6), axis-right
length graphic, and Calculate storing overlayPng without touching batchApply[cal].
After calibration Calculate, reuse approved sheet/sensitivity on non-calibration
photos via the mask worker + inventory compose. Persist per-photo ok/error on
measure.batchApply and the device draft for #64; leave Mask UI unopened and keep
scale/widths/sqft/bdft on the calibration results only.
Document single-source M1 path after pickCalibration: Mask/Axis/Measure/
Calculate use the chosen Source only; batch siblings stay draft-untouched
until #62 apply. Tag encode-fallback inventory Blob as image/png (CAPTURE.md).
Proving tests: after pick of index N≠0, Mask→Axis→Measure operate on that
blob only; Calculate yields inventory PNG + M1 numbers; other batch photos
stay in draft.batch untouched; resume through measure restores calibration
source. FakeWorker drives the path in Node. No production change yet (TDD).
Failing session/UI tests for M2-3: Continue → pickCalibration (not
first-photo Mask proxy), user pick enters Mask on that Source, draft
remembers calibrationIndex, Resume restores. No production change yet (TDD).
Replace single Add photo with Batch: accumulate 1–N photos, Continue
when ≥1 (M1 bridge: mask on first), persist batch blobs in the device
draft, Resume Continue restores. Keeps Mask/Axis/Measure intact.
Failing session/UI tests for M2-2: multi-photo batch state, Continue
when ≥1, device-draft persistence of batch blobs, Resume Continue
restores. No production change yet (TDD).
Lock Ty’s decisions: width stations start 6in from EXIF-oriented
photo bottom along the approved length axis; lineup is clean cal,
dimension overlay, then other batch PNGs in capture order.
sampleWidths used floor(L/6)+1 stations from t=0, so tapered tips
dominated min/max. Sample only interior stations at 6, 12, … ≤ L−6.
When L≤12 there is no interior span: every6In stays empty and
min/max/avg stay 0. CAPTURE.md Widths bullet matches.
sqft, bdft, scale, and inventory PNG are unchanged.
Ty UAT: width min/max must describe bulk, not tips. Pin L=48 tapered
case (7 interior stations, no tip endpoints), uniform mid-board count,
and L≤12 empty/zero fallback before the sampleWidths change lands.
Product call for the WooCommerce redo. Calculate accepts
sources whose post-crop shorter side is at least 800px.
Warn copy and CAPTURE.md match. No upscale still holds.
Ty UAT rejected phone photos under the 1600px bar.
Lock the mid-band case (shorter side 1000) and floor
geometry at 800×1067 / 0.8×800=640 before the code change.
Why
CAPTURE.md Crop keeps slab pixels and makes removed pixels transparent.
encodeInventoryPng only drawImage-cropped the AABB from the raw photo.
Scope
composeInventoryFrame samples the source through approvedMask.bitset.
Non-slab pixels stay alpha 0. Drop the 8% MARGIN; fill 80% of mask
extents in a 3:4 frame. Floor rules unchanged.
Verification
cd frontend && npm test
12 pass, including alpha contract (no sheet green in opaque pixels).
Why
Calculate showed accurate numbers but the inventory PNG kept the sheet
background. Measure already uses approvedMask.bitset; encode did not.
Scope
Add composeInventoryFrame and inventory.test.ts alpha/crop proofs.
Wire encodeInventoryPng({ bitmap, mask, source }) at calculate().
This commit still paints the slabBounds AABB without the bitset gate so
the alpha contract fails until the follow-up fix.
Verification
node --import tsx --test src/png/inventory.test.ts
alpha contract fails (sheet green leaks); 3:4 / 80% / floor pass.
Replace MeasurePanel steppers with number inputs. Commit on blur, change,
and Calculate snaps inches to nearest 1/8 via inToEighths. Calculate
enables when both drafts parse to > 0. Axis, mask, and scale math unchanged.
Ty
grokbot_ty
grokbot_ty