recognizer Leaf 7 Core side: recoder SetupDecoder beam maps (7a) + RecodedCharId::from_codes (7b) + board by AdaWorldAPI · Pull Request #647 · AdaWorldAPI/lance-graph

AdaWorldAPI · 2026-07-04T18:12:58Z

The lance-graph (OGAR Core) side of the Tesseract recognizer's Leaf 7 (the CTC decode). The compute-side beam lands in tesseract-rs PR #6; this PR is the two Core surfaces it consumes plus the board findings. Every leaf is byte-parity-proven against a live libtesseract 5.3.4 oracle.

Leaf 7a — recoder `SetupDecoder` beam maps (`3242f6bb`, CODE)

The deferred half of the recoder leaf (E-CPP-PARITY-7 built only decoder_). setup_decoder now transcodes the full C++ SetupDecoder (unicharcompress.cpp:395-436): the decoder_ map and the three beam-search trie maps the recognizer walks — is_valid_start_ (IsValidFirstCode), final_codes_ (GetFinalCodes), next_codes_ (GetNextCodes), surfaced as public accessors. RecodedCharId gains code_at + truncated. Byte-parity GREEN on real eng.lstm-recoder: dump_beam diffs 114 lines byte-identical vs a new beam mode in recoder_oracle.cpp (deterministic walk, since unordered_map order is unspecified). Real data corrected the docs: code_range=111, one shared code (id1/id2→110), final_codes_[<empty>] push order 0,110,1…109, next_codes_ empty (all length-1). Hand-traced length-3 unit test exercises the full trie. E-OCR-RECODER-BEAM-1.

Leaf 7b dependency — `RecodedCharId::from_codes` (`26370b0e`, CODE)

The public from_codes(&[i32]) constructor (the C++ RecodedCharID::Set loop) — the beam's key builder: it makes a prefix (codes[0..length]) to query get_final_codes/get_next_codes and a full_code (prefix ++ code) to feed decode. The last Core surface the beam needs.

Board

E-OCR-RECODER-BEAM-1 (7a) + E-OCR-RECODEBEAM-1 (7b — the non-dict CTC beam, byte-parity green across 4 configs vs libtesseract; the "recognizer produces text" milestone) prepended to EPIPHANIES.

+2 tests (815 contract total); clippy -D warnings + fmt clean (-p lance-graph-contract scoped). Branch restarted from main after #643/#644/#645 merged.

Merge order

Merge this before tesseract-rs PR #6 — that PR's lance-graph-contract path dep builds against lance-graph main, and its beam needs both the 7a accessors and from_codes. Expect #6 CI red until this merges.

🤖 Generated with Claude Code

https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1

…reen The recurrent layer, the hardest recognizer leaf: 4 gates via Leaf 4 (fully_connected_forward) + the cell recurrence c=clip(GF1·c+CI·GI,±100), output h=tanh(c)·GO, and the int8-quantized recurrent feedback. Byte-parity green across 3 shapes incl. ns=48/ni=36 × 8 timesteps vs a libtesseract oracle running the REAL per-timestep LSTM::Forward body. Code in tesseract-rs #5 (recognizer Leaves 1-5); board hygiene lands here per the CLAUDE.md rule. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1

… green The composition (graph::Layer: Series/Reversed/Parallel) that chains the proven layer leaves into a network forward, with the inter-layer int8 requant. Byte-parity green: Series[LSTM,FC] across 4 shapes incl. ns=96/ni=192/no=111 (eng.lstm's LSTM192→Fc111 tail) vs a libtesseract oracle. Code in tesseract-rs #5 (Leaves 1-6); board lands here per the CLAUDE.md rule. Next: Leaf 7 recodebeam (CTC) → text. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1

coderabbitai · 2026-07-04T18:13:15Z

📝 Walkthrough

Walkthrough

This PR adds two documentation entries to the EPIPHANIES.md board file, recording byte-parity validation findings for recognizer pipeline components: Leaf 6 (graph walk composition) and Leaf 5 (LSTM::Forward int8), both marked as GREEN parity against a libtesseract oracle.

Changes

Epiphany log updates

Layer / File(s)	Summary
Add Leaf 6 and Leaf 5 epiphany entries `.claude/board/EPIPHANIES.md`	Documents byte-parity GREEN results for graph-walk composition (Series/Reversed/Parallel) with int8 requantization, and for LSTM::Forward 1-D int8 per-timestep recurrence, including test notes and cross-references.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Poem

A rabbit hops through boards of green,
Two findings logged, both crisp and clean.
Byte for byte, the parity holds,
LSTM whispers what the graph walk told.
🐇✨ Onward to the next epiphany scene!

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Title check	⚠️ Warning	The title points to Leaf 7 recoder changes, but the PR only updates the board with findings for Leaves 5 and 6.	Rename it to reflect the board-only update and the actual Leaf 5/6 byte-parity findings.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ Finishing Touches

📝 Generate docstrings

Create stacked PR
Commit on current branch

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

…-parity green Transcode the full C++ UnicharCompress::SetupDecoder (unicharcompress.cpp:395-436) into the Core's UnicharCompress — the deferred half of the recoder leaf. In one ascending-id pass, setup_decoder now builds the decoder_ map AND the three beam-search trie maps that RecodeBeamSearch consumes: - is_valid_start_ (Vec<bool>): code(0) is a valid first code (IsValidFirstCode) - final_codes_ (prefix -> completing last codes, GetFinalCodes) - next_codes_ (prefix -> non-final continuations, GetNextCodes) The `while (--len >= 0)` prefix walk climbs from the direct parent toward the empty prefix, deduping and breaking at the first already-seeded next prefix. RecodedCharId gains code_at (C++ operator()) + a private truncated (C++ Truncate; trailing code slots drop out of identity since eq/hash read code[0..length]). Three new public accessors surface the maps read-only for the recognizer's beam search (Leaf 7b): is_valid_first_code / get_final_codes / get_next_codes. Per Core-First the table is Core content; the beam SEARCH that walks it is recognizer compute. Byte-parity GREEN on real eng.lstm-recoder: dump_beam (a deterministic walk, since unordered_map order is unspecified) diffs 114 lines byte-identical vs the new `beam` mode in recoder_oracle.cpp. Real data corrected the module docs: code_range=111 (max code 110, not 112), one shared code (id1/id2 -> 110), final_codes_[<empty>] push order 0,110,1,2..109; next_codes_ empty (all length-1). A hand-traced length-3 unit test exercises the full trie. encode + decode regressions still green. +2 tests (814 contract total); clippy -D warnings + fmt clean (-p lance-graph-contract scoped). Cross-ref E-OCR-RECODER-BEAM-1, E-CPP-PARITY-7, E-OCR-GRAPHWALK-1. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1

… (Leaf 7b) Add the public `RecodedCharId::from_codes(&[i32])` constructor (the C++ `RecodedCharID::Set` loop, unicharcompress.h:43) — the recognizer's RecodeBeamSearch (Leaf 7b, in tesseract-core) builds `prefix` (codes[0..length]) to query get_final_codes/get_next_codes and `full_code` (prefix ++ code) to feed DecodeUnichar. This is the last Core surface the beam needs; the beam itself (byte-parity green across 4 configs vs libtesseract on eng.lstm-recoder) lands in tesseract-rs PR #5. Board: prepend E-OCR-RECODEBEAM-1 (Leaf 7b — the non-dict CTC beam, the "recognizer produces text" milestone). +1 test (815 contract total); clippy -D warnings + fmt clean (-p lance-graph-contract scoped). Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016b33swuXE23hKtqxsHu9p1

claude added 2 commits July 4, 2026 17:23

AdaWorldAPI changed the title ~~board: recognizer Leaves 5-6 (LSTM::Forward + graph walk) byte-parity findings~~ recognizer: Leaf 7a recoder SetupDecoder beam maps (Core code) + Leaves 5-6 board findings Jul 4, 2026

AdaWorldAPI changed the title ~~recognizer: Leaf 7a recoder SetupDecoder beam maps (Core code) + Leaves 5-6 board findings~~ recognizer Leaf 7 Core side: recoder SetupDecoder beam maps (7a) + RecodedCharId::from_codes (7b) + board Jul 4, 2026

AdaWorldAPI merged commit b74378c into main Jul 4, 2026
6 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

recognizer Leaf 7 Core side: recoder SetupDecoder beam maps (7a) + RecodedCharId::from_codes (7b) + board#647

recognizer Leaf 7 Core side: recoder SetupDecoder beam maps (7a) + RecodedCharId::from_codes (7b) + board#647
AdaWorldAPI merged 4 commits into
mainfrom
claude/happy-hamilton-0azlw4

AdaWorldAPI commented Jul 4, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jul 4, 2026 •

edited

Loading

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AdaWorldAPI commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Leaf 7a — recoder SetupDecoder beam maps (3242f6bb, CODE)

Leaf 7b dependency — RecodedCharId::from_codes (26370b0e, CODE)

Board

Merge order

Uh oh!

coderabbitai Bot commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

AdaWorldAPI commented Jul 4, 2026 •

edited

Loading

Leaf 7a — recoder `SetupDecoder` beam maps (`3242f6bb`, CODE)

Leaf 7b dependency — `RecodedCharId::from_codes` (`26370b0e`, CODE)

coderabbitai Bot commented Jul 4, 2026 •

edited

Loading