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

.claude/board/EPIPHANIES.md

@@ -1,3 +1,42 @@
+## 2026-07-04 — E-OCR-RECODEBEAM-1 — recognizer Leaf 7b: `RecodeBeamSearch::Decode` (the non-dict CTC beam) is byte-parity green — logits → labels → text, the "recognizer produces text" milestone (the hardest leaf, 1382 lines)
+**Status:** FINDING (byte-parity proven vs libtesseract 5.3.4; `tesseract-core`, tested)
+
+The CTC beam decode ships — the single hardest recognizer leaf (`recodebeam.cpp`, 1382 lines), the step that turns the LSTM's per-timestep softmax logits into a code sequence → text. `tesseract_core::recodebeam::RecodeBeamSearch` transcodes the **non-dictionary** path (`dict_ == nullptr`, `permuter` fixed at `TOP_CHOICE_PERM`, the dawg beams never populated). Placed in **tesseract-core** (not `tesseract-recognizer`): the beam is compute but **SIMD-free and recoder-coupled**, so it belongs next to the recoder tables it walks (Leaf 7a's `is_valid_first_code`/`get_final_codes`/`get_next_codes`) and the `recoded_to_text` (`E-CPP-PARITY-7`) step it feeds. Transcoded whole: `ComputeTopN` (top-n flags via the min-heap), `DecodeStep` (per-timestep, top-n-group fallback), `ContinueContext` (the beam-extension crux — prefix walk + `get_final_codes`/`get_next_codes` continuations + the dup/null CTC combination rules), `ContinueUnichar`, `PushDupOrNoDawgIfBetter`, `PushHeapIfBetter` (`score = cert + prev.score`, evict-worst-when-full), `ComputeCodeHash` (the u64 rolling mix for dup-path removal), `UpdateHeapIfMatched` (+`Reshuffle`), `ExtractBestPaths`/`ExtractBestPathAsLabels`, plus a faithful `GenericHeap<KDPairInc<f64,…>>` (binary min-heap, exact `SiftUp`/`SiftDown`/`Reshuffle` so the `get(i)` internal order the decode walks matches). The C++ borrowed-`prev`-pointer lattice becomes a safe **arena** (`Vec<RecodeNode>` + `prev: Option<u32>` index) — no `unsafe`, no dangling across heap reshuffles.
+
+Byte-parity **GREEN** on the real `eng.lstm-recoder`: the public `extract_best_path_as_labels` diffs **byte-identical** across **4 configs** — `null_char ∈ {110, 0, 42}` × {fold, simple-text} — vs a libtesseract oracle (`/tmp/recodebeam_oracle.cpp`) that constructs the REAL `RecodeBeamSearch(recoder, null_char, simple, nullptr)` and runs the public `Decode(GENERIC_2D_ARRAY<float>, 1.0, 0.0, 0.0, nullptr)` + `ExtractBestPathAsLabels` on the SAME synthetic softmax matrix (read from a shared `.bin`, so the input is byte-identical). Using only the **public** ctor + Decode + ExtractBestPathAsLabels means **no private-member access** → the 5.5.0-header / 5.3.4-lib ABI skew that dogged the earlier leaves cannot bite. The folding path proved `[5,5,·,7,7,7,·,9] → [5,7,9]` (drop nulls, fold adjacent); the simple-text path proved the no-fold branch. **Float contract:** scores accumulate in f32 (`TFloat=float`), `ProbToCertainty(p) = p>e^-20 ? ln(p) : -20` is a raw `ln` (NOT the Leaf-3 LUT), heap keys are `f64::from(score)` (lossless → order-preserving) — no FP drift vs libtesseract. **Core addition:** `RecodedCharId::from_codes` (the beam's prefix/full-code key builder, `unicharcompress.h:43` `Set` loop). eng.lstm's recoder is pass-through (all length-1) so `next_codes_` is empty and every beam sits at length 0 — this proves the CTC core for a simple (non-CJK) script; the multi-code `next_codes_` trie is Han/Hangul, out of `eng` scope (consistent with every prior leaf, and the `next_codes_` maps themselves are proven in 7a). +4 unit tests (12 `tesseract-core` total); clippy `-D warnings` + fmt clean (`-p tesseract-core` scoped).
+
+**The recognizer now spans logits → text:** the 1-D forward (Leaves 1-6) produces the softmax logits, and Leaf 7 (7a maps + 7b beam) decodes them → codes → `recoded_to_text` → the string. The remaining gap to **image → text** is the **2-D front-end** (`Convolve`/`Maxpool`/`Reconfig`/`XYTranspose` + the `NetworkIO`/`StrideMap` grid + the leptonica image `Input`); the dictionary / language-model beam is the later accuracy layer. Cross-ref: `E-OCR-RECODER-BEAM-1` (7a, the maps this walks), `E-CPP-PARITY-7` (`recoded_to_text`, the text step it feeds), `E-OCR-GRAPHWALK-1` (Leaf 6, the logits it decodes). Plan `tesseract-rs/.claude/plans/recognizer-decode-frontend-v1.md` (Leaf 7a + 7b EXECUTED). lance-graph PR #647 (`from_codes` + board); tesseract-rs PR #5 (the beam). Branch `claude/happy-hamilton-0azlw4`.
+
+## 2026-07-04 — E-OCR-RECODER-BEAM-1 — recognizer Leaf 7a: the recoder `SetupDecoder` beam-search trie maps (`is_valid_start_`/`final_codes_`/`next_codes_`) are byte-parity green — the deferred half of the recoder leaf, the surface `RecodeBeamSearch` walks
+**Status:** FINDING (byte-parity proven vs libtesseract 5.3.4; `lance-graph-contract`, tested)
+
+The deferred half of the recoder leaf (`E-CPP-PARITY-7` built only the `decoder_` map) now ships in the Core's `UnicharCompress`. `setup_decoder` transcodes the **full** C++ `SetupDecoder` (`unicharcompress.cpp:395-436`) in one ascending-id pass: the `decoder_` map **and** the three beam-search trie maps — `is_valid_start_` (`code(0)` is a valid first code, `IsValidFirstCode`), `final_codes_` (prefix `code[0..len-1]` → the 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 stopping at the first already-seeded `next_codes_` entry. `RecodedCharId` gained `code_at` (C++ `operator()`) + a private `truncated` (C++ `Truncate`, which only sets `length_` — the trailing `code` slots drop out of identity since eq/hash read `code[0..length]`). These maps are **Core content** (computed table state loaded alongside the encoder, data-shaped, no lifecycle); the recognizer's `RecodeBeamSearch` (Leaf 7b) *consumes* them read-only via the three new public accessors `is_valid_first_code`/`get_final_codes`/`get_next_codes`. Per Core-First: the table is Core, the beam **search** that walks it is recognizer compute.
+
+Byte-parity **GREEN** on the real `eng.lstm-recoder`: `dump_beam` (a deterministic walk — `unordered_map` iteration order is unspecified and differs from Rust's `HashMap`, so the dump drives off `encoder_` id-order × truncation-length ascending, deduped, querying the accessors) diffs **114 lines byte-identical** vs the `beam` mode added to `recoder_oracle.cpp` (links libtesseract, same deterministic walk via `IsValidFirstCode`/`GetFinalCodes`/`GetNextCodes`). The real data corrected an assumption baked into the module docs: `code_range = 111` (max code 110, **not** 112), `enc_size = 112`, one shared code (`roundtrip_bad=1`, the id1/id2 → code 110 case), and `final_codes_[<empty>]` in push order `0,110,1,2,…,109` (id1 seeds 110, id2 dedups, then codes 1..109). `next_codes_` is empty (all length-1 → the `--len` loop never runs); a hand-traced length-3 (Han/Hangul-shaped) unit test exercises the full trie (the `--len` climb, the dedup-then-`break`, multiple finals under one prefix). encode + decode regressions still green. +2 unit tests (814 contract tests total); clippy `-D warnings` + fmt clean (`-p lance-graph-contract` scoped).
+
+**Leaf 7a unblocks Leaf 7b** — `RecodeBeamSearch::Decode` (the CTC beam search, `recodebeam.cpp` 1382 lines): `ComputeTopN` per-timestep → `DecodeStep`/`ContinueContext` walking these `next_codes_`/`final_codes_`/`is_valid_start_` maps → `ExtractBestPaths` → the code lattice → `recoded_to_text` (`E-CPP-PARITY-7`) → the string. Non-dict first pass (the falsifiable core; dict / language-model beam deferred). Cross-ref: `E-CPP-PARITY-7` (the recoder load side + `recoded_to_text`), `E-OCR-GRAPHWALK-1` (Leaf 6, produces the softmax logits the beam decodes), plan `tesseract-rs/.claude/plans/recognizer-decode-frontend-v1.md` (Leaf 7a EXECUTED). lance-graph PR #647 (recognizer Core side); tesseract-rs PR #5 (Leaves 1-6). Branch `claude/happy-hamilton-0azlw4`.
+
+## 2026-07-04 — E-OCR-GRAPHWALK-1 — recognizer Leaf 6: the graph walk (`Series`/`Reversed`/`Parallel`) is byte-parity green — the composition that chains the proven layer leaves into a network forward, with the inter-layer int8 requant
+**Status:** FINDING (byte-parity proven vs libtesseract 5.3.4; `tesseract-recognizer`, tested)
+
+The composition that turns the individual layer leaves into a network forward — the compute-side **execution tree**, the `invoke_network` counterpart (the Core's `lance_graph_contract::network` FacetCascade describes the tree *structure*; `tesseract_recognizer::graph::Layer` *runs* it). NOT a parallel object model: it is the runnable subset (the layers whose `Forward` this crate transcodes), built from the Core's tree by a consumer. `Layer { Lstm, FullyConnected, Reversed, Series, Parallel }`:
+- **`Series`** (`series.cpp:Forward`): run each sub-layer in turn, output of N → input of N+1. The recognizer runs int8, so the intermediate `NetworkScratch::IO` buffers inherit `int_mode` → the inter-layer conversion is the **int8 requant** (`quantize_i8` = `NetworkIO::WriteTimeStep`, proven Leaf 5). The final softmax is the `ResizeFloat` exception.
+- **`Reversed`** (XREVERSED, `reversed.cpp`): reverse the 1-D sequence → inner → reverse. (`Txy`/`YREVERSED` are 2-D front-end, deferred.)
+- **`Parallel`** (`parallel.cpp`): same input to each sub-layer, concatenate outputs.
+
+Byte-parity **GREEN**: `Series[LSTM, FC(tanh)]` across 4 shapes incl. `ns=96/ni=192/no=111` (eng.lstm's LSTM192→Fc111 tail) — diffing f32 bit-patterns vs a libtesseract oracle (`/tmp/graph_oracle.cpp`, `-DFAST_FLOAT`) that runs the REAL LSTM per-timestep body over all timesteps → the REAL `WriteTimeStep` int8 requant → the REAL `MatrixDotVector`+`FuncInplace` — the exact `Series::Forward` composition. Proves the chaining order + the inter-layer requant (the one debug: the SIMD `MatrixDotVector` over-writes `RoundOutputs` padding, so oracle output vectors need `+128`). +4 unit tests (27 total); clippy `-D warnings` + fmt clean.
+
+The recognizer now chains the proven layers into a **1-D network forward** from feature sequences → softmax logits. **Next Leaf 7 = `recodebeam`** — the CTC beam decode (needs the recoder's deferred `SetupDecoder` beam maps: `GetNextCodes`/`GetFinalCodes`/`is_valid_start_`) → the code lattice `recoded_to_text` (E-CPP-PARITY-7) eats → the text string. Then the 2-D front-end (`Convolve`/`Maxpool`/`XYTranspose` + the `NetworkIO`/`StrideMap` grid + leptonica image `Input`) closes image→text. Cross-ref: `E-OCR-LSTM-1` (Leaf 5), `E-OCR-FULLYCONNECTED-1` (Leaf 4), `E-CPP-PARITY-7` (the recoder `recoded_to_text` the lattice feeds). tesseract-rs PR #5 (recognizer Leaves 1-6). Branch `claude/happy-hamilton-0azlw4`.
+
+## 2026-07-04 — E-OCR-LSTM-1 — recognizer Leaf 5: `LSTM::Forward` (1-D int8) is byte-parity green vs libtesseract — the recurrent layer, the hardest leaf, composes 4 Leaf-4 gates + the int8-quantized state recurrence
+**Status:** FINDING (byte-parity proven vs libtesseract 5.3.4; `tesseract-recognizer`, tested)
+
+The recurrent layer ships — the hardest recognizer leaf. `tesseract_recognizer::Lstm` transcodes the 1-D non-softmax `LSTM` (`lstm/lstm.{h,cpp}`, the eng.lstm case). `from_le_bytes` = `LSTM::DeSerialize` (`lstm.cpp:253-287`): `i32 na_` + 4 gate `WeightMatrix`es (`CI, GI, GF1, GO`, `GFS` skipped for 1-D), `ns = CI.num_outputs`, `ni = na_ − ns`. `forward` = `LSTM::Forward` (`lstm.cpp:363-454`) per timestep: source = `[input | int8_quantize(prev_output)]`; the 4 gates via **`fully_connected_forward` (Leaf 4)** — CI=tanh(`GFunc`), GI/GF1/GO=logistic(`FFunc`); cell `c = clip(GF1·c + CI·GI, ±100)` (`kStateClip`); output `h = tanh(c)·GO` (`HFunc`). The crux of the int8 path: the recurrent `h` is **quantized back to int8** (`NetworkIO::WriteTimeStepPart`, `networkio.cpp:662-666`: `clip(IntCastRounded(x·127), ±127)`, round-half-away-from-zero, never −128) before the next timestep's gate matmuls.
+
+Byte-parity **GREEN** across 3 shapes incl. `ns=48/ni=36` (eng.lstm's LSTM48) × **8 timesteps** — diffing f32 bit-patterns vs a libtesseract oracle (`/tmp/lstm_oracle.cpp`, `-DFAST_FLOAT`) that runs the REAL `WeightMatrix::MatrixDotVector` + `FuncInplace<GFunc/FFunc>` + `MultiplyVectorsInPlace`/`MultiplyAccumulate`/`ClipVector` + `FuncMultiply<HFunc>` + the real `IntCastRounded`/`ClipToRange` quant — the exact per-timestep body of `LSTM::Forward`. The 8-timestep run proves the int8 recurrence feedback is correct across the sequence. **No FMA discrepancy** — the recurrence's separate mul+add matches libtesseract byte-exactly at `-O2`. Added `WeightMatrix::from_le_bytes_prefix` (returns bytes consumed) to chain the 4 gate matrices. +5 unit tests (23 total); clippy `-D warnings` + fmt clean (`-p tesseract-recognizer` scoped).
+
+The recognizer now has all four leaf types of the eng.lstm forward pass: int8 GEMM (Leaf 1), WeightMatrix (Leaf 2), activations (Leaf 3), FullyConnected layer (Leaf 4), LSTM layer (Leaf 5). **Next Leaf 6 = the graph walk** — `Series`/`Parallel`/`Reversed`/`Convolve`/`Maxpool` composing the layers per the model tree (`[1,36,0,1[C3,3Ft16]Mp3,3TxyLfys48Lfx96RxLrx96Lfx192Fc111]`) → then `recodebeam` (CTC decode → the code lattice `recoded_to_text` eats). Cross-ref: `E-OCR-FULLYCONNECTED-1` (Leaf 4, reused per gate), `E-OCR-ACTIVATION-1` (Leaf 3), `E-OCR-WEIGHTMATRIX-1` (Leaf 2), `E-OCR-NETWORK-SINK-1` (the structure side). Plan: `tesseract-rs/.claude/plans/recognizer-core-shape-v1.md` (Leaf 5 EXECUTED). tesseract-rs PR #5 (recognizer Leaves 1-5). Branch `claude/happy-hamilton-0azlw4`.
+
 ## 2026-07-04 — E-V1-TAIL-FORBIDDEN-V3-IS-CONTENT-BLIND-1: the flat contiguous-u24 tail (`family:u24 ++ identity:u24`) is V1-LEGACY, FORBIDDEN for new units; V3 is the content-blind 4+12 facet whose 12B the ClassView projects as axis-grouped byte rails — migration of residues MANDATORY
 **Status:** RULING (operator, 2026-07-04). Specializes/enforces `E-V3-FACET-4-PLUS-12` (operator-locked 2026-07-02) with an explicit *forbidding* + a *mandatory-migration* stamp. Reverses the older 2026-06-10 OGAR P0 pin (u24+u24) and this repo's own 2026-06-13 CANON block (which still document the flat tail) — latest operator word wins; both are regraded in place, append-only, never deleted.
 

crates/lance-graph-contract/examples/recoder_dump.rs

@@ -17,6 +17,12 @@
 //! diff /tmp/oracle_recoder_encode.tsv /tmp/rust_recoder_encode.tsv \
 //!   && diff /tmp/oracle_recoder_decode.tsv /tmp/rust_recoder_decode.tsv
 //! # both byte-identical => the recoder load-side is byte-parity green
+//!
+//! # The `beam` mode dumps the SetupDecoder trie maps (is_valid_start_ /
+//! # final_codes_ / next_codes_ — the RecodeBeamSearch surface, Leaf 7a):
+//! #   ./recoder_oracle /tmp/eng.lstm-unicharset /tmp/eng.lstm-recoder beam > /tmp/oracle_recoder_beam.tsv
+//! cargo run -p lance-graph-contract --example recoder_dump -- /tmp/eng.lstm-recoder beam > /tmp/rust_recoder_beam.tsv
+//! diff /tmp/oracle_recoder_beam.tsv /tmp/rust_recoder_beam.tsv   # byte-identical => 7a green
 //! ```
 
 #![allow(
@@ -39,6 +45,7 @@ fn main() -> ExitCode {
         Ok(recoder) => {
             match mode.as_str() {
                 "decode" => print!("{}", recoder.dump_decode()),
+                "beam" => print!("{}", recoder.dump_beam()),
                 _ => print!("{}", recoder.dump_encode()),
             }
             ExitCode::SUCCESS