Skip to content

board: E-OCR-MAKEROW-2 + -3 — makerow row-assignment/cleanup + x-height byte-parity findings #659

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AdaWorldAPI merged 2 commits into from
Jul 7, 2026
+18 −0
Merged

board: E-OCR-MAKEROW-2 + -3 — makerow row-assignment/cleanup + x-height byte-parity findings #659

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6775608
board: E-OCR-MAKEROW-2 — 3E wave 2 byte-parity 4/4 + the tie-order pa…
claude Jul 7, 2026
eac238e
board: E-OCR-MAKEROW-3 — 3E wave 3 x-height chain byte-parity 4/4; ru…
claude Jul 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 18 additions & 0 deletions .claude/board/EPIPHANIES.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,21 @@
## 2026-07-07 — E-OCR-MAKEROW-3 — 3E wave 3: the x-height chain byte-parity 4/4 — the fixed ruff (#57) made the `Textord::` methods harvestable first-class, closing the make_rows textline stage
**Status:** FINDING (byte-parity proven vs source-compiled C++ x-height chain; `tesseract-ocr`, tested)

The x-height stage — `Textord::compute_block_xheight` / `Textord::compute_row_xheight` (the two out-of-line class methods that were INVISIBLE to `walk_free_functions` until ruff #57 landed the class-qualified capture; the refreshed manifest `makerow-callgraph-v3.txt` shows their full dispatch edges) + `compute_xheight_from_modes`, `compute_row_descdrop`, `correct_row_xheight`, and the `makerow.h` inlines (`get_min_max_xheight`/`get_row_category`/`within_error_margin`) — lands on waves 1-2 (tesseract-rs `4faae42`). `ToRow` gains xheight/ascrise/descdrop/evidence/all_caps/rep_chars_marked; `ToBlockCtx` gains xheight. Per-subexpression precision transcribed (`median()`→f32 narrowing, asc_frac 0.5f, `kXHeightCapRatio` double).

**Byte-parity 4/4** (2 seeds × clean/noise) vs the wave-2 oracle extended with the x-height chain transcribed VERBATIM against the real compiled `STATS`: the new STAGE7 (per-row xheight/ascrise/descdrop/category + block xheight) is byte-identical AND every wave-2 STAGE1-6 prefix line is unchanged (the additive-only bar held). PARITY PINs (per E-OCR-MAKEROW-2): `mark_repeated_chars` reduces to a documented no-op on the plain-tuple carrier; `joined_to_prev` always false. No new sort/nth_element → no new tie-order pin. 68/68 lib tests (5 new). **The ruff→port loop closed the circle:** the harvest gap fixed in ruff #57 was the exact prerequisite for this wave — the Textord:: methods it exposed ARE this wave's port targets.

**The make_rows textline formation is now COMPLETE through x-height** (waves 1+2+3: STATS/DetLineFit/occupation math → row assignment + cleanup → x-height). Remaining before the parity line finder replaces `seg-approx`: the straight-baseline hookup (`textord_straight_baselines` path — the parallel fit already computed) + the make_rows→line-box wiring (3F₁ PSM_SINGLE_LINE, 3F₂ PSM_SINGLE_BLOCK). Then P6 golden corpus. Note: run INLINE on the main thread (both delegated wave-3 agents were stopped mid-run). Plan `tesseract-rs/.claude/plans/pdf-to-text-ocr-v1.md`. Branch `claude/happy-hamilton-0azlw4`.

## 2026-07-07 — E-OCR-MAKEROW-2 — 3E wave 2: the makerow row-assignment + cleanup chain byte-parity 4/4 — and the TIE-ORDER finding (unspecified in real Tesseract, pinned in the probe)
**Status:** FINDING (byte-parity proven vs source-compiled C++ leaves; `tesseract-ocr`, tested)

The chain `assign_blobs_to_rows`/`most_overlapping_row` (pass semantics, skew interpolation, the merge branch's iterator bookkeeping mapped to indices) → `fit_parallel_rows`/`fit_parallel_lms` (wave-1 ConstrainedFit) → `delete_non_dropout_rows` (the proven occupation leaves wired into the real prune decision) → `expand_rows`/`adjust_row_limits`/`compute_row_stats` → `make_rows` + REAL `compute_page_skew` lands on the wave-1 leaves (tesseract-rs `ad71e01`), with `ToRow`/`ToBlockCtx` as documented working-set carriers (TO_ROW ctor/add_blob/set_limits verbatim, `blobbox.cpp:685-766`). **Byte-parity 4/4** (2 seeds × clean/noise, per-stage f32-hex dumps) vs an oracle compiling `statistc.cpp`/`detlinefit.cpp` from source.

**The transferable finding:** real Tesseract's `blob_x_order` sort and `row_spacing_order` `nth_element` leave TIE ORDER UNSPECIFIED (qsort/introselect semantics) — and `TO_ROW::add_blob`'s expansion is ORDER-DEPENDENT, so two faithful shells can diverge on equal-key inputs while both being "correct". The probe pins identical total orders on BOTH sides (full box tuple; spacing→intercept→min_y), documented in-code as PARITY PINs. Any future textord leaf with a sort/nth_element must get the same treatment up front. (Wave note: the porting agent was stopped mid-run; the orchestrator finished from its working tree — state was 90% complete and correct, the last 10% was exactly this tie-order class.)

Next: 3E wave 3 (the xheight chain — `compute_block_xheight`/`compute_row_xheight`/`correct_row_xheight` + straight-baseline fallback), now harvestable FIRST-CLASS: ruff #57 (merged) made the out-of-line `Textord::` methods visible with class-qualified dispatch edges. Then the parity line finder replaces `seg-approx`, then P6. Plan `tesseract-rs/.claude/plans/pdf-to-text-ocr-v1.md`. Branch `claude/happy-hamilton-0azlw4`.

## 2026-07-07 — E-HOTPLUG-MIGRATION-1 — the generic hot-plug pattern is the migration target for ALL consumers; migration doc is board discipline
**Status:** RULING (operator, 2026-07-07; shipped OGAR #174/#175 + lance-graph #658 + tesseract-rs #13/#14)

Expand Down