The byteflow lane, round 3 — the dynamic coding-rate patcher and its wiring

Lane: wt/byteflow3 (base origin/wt/byteflow2, the resurrected crate), 2026-10-02. The mission: the owner’s decision — «срочно внедрить наш byteflow», с патчером по coding-rate — решение владельца, не BLT-энтропия, не фиксированный шаг:

• R_ε(H) = ½·logdet(I + (d/ε²)·H·Hᵀ) on the hidden states — paper eq. (11);
• ΔR_t = R_ε(h_1:t) − R_ε(h_1:t−1) — the marginal contribution of byte t, eq. (12);
• patch borders = Top-K by ΔR_t, BOS forced, chronological — §3.2’s own selection rule, the paper’s argument against a global threshold being that it breaks the static compute graph.

1. What the paper’s recipe is (and what was NOT invented)

The chunker scores the local encoder’s hidden states h_1:T ∈ R^{T×d_local} (p.4 §3.2: “the local encoder produce contextualized representations”). That is the «лёгкий драфт-модуль»: SWA + Canon layers, linear cost — not the main model. The crate (vendor/dormouse-fused/crates/burn-byteflow) already carried this against the byteflow2 findings doc; what was missing is the piece that lets it sit behind a BYTE STREAM (dormouse-data streams bytes, window W) — the patcher as STATE:

• burn_byteflow::RatePatcher (src/stream.rs): push(h: &[f32]) -> ΔR_t per position; window state is the d×d Gram (LogDet) or one running sum of squared norms (L2) — O(d²) memory, never a T-row buffer, so a stream of any length patches in bounded memory; reset() begins the next window (one window = one sequence h_1:W, the paper’s own per-sequence Top-K semantics — carrying the Gram across windows would change what ΔR means).
• The two modes are the paper’s own pair: LogDet = eq. (11) computed incrementally through the same Cholesky + Sylvester route the batch path takes (ponytail:-marked O(d³) per push; a maintained rank-1 Cholesky update is the O(d²) upgrade if a profile ever asks); L2 = Appendix B’s own streaming approximation (R ∝ ‖H‖₂), O(d) per push.
• borders(k) = {position 0} ∪ top-(K−1) by ΔR, chronological, ties broken on position — the same set select_positions returns from batch gains.

2. The oracle — formula against the direct log-det, green

tests/dynamic_patcher_oracle.rs (CPU ndarray, f64 on both sides):

gate	what it pins	bar
test_coding_rate_matches_direct_logdet	the streamed rate AND every prefix ΔR_t against the paper’s LITERAL T×T eq. (11) matrix I_T + (d/ε²)·H·Hᵀ, log-det taken by LU with partial pivoting — a third decomposition of the same equation (no Cholesky, no eigen, no Sylvester); plus the telescope identity (ΣΔR = R(h_1:T)) and the d×d direct cross-check that the Sylvester folding is right	1e-9 rel (measured: far below)
test_deltaR_topk_borders	a 64-position stream with three planted 100× bursts at 10/30/50: the borders land on the bursts in BOTH modes, BOS stays first, chronological, borders(1)=[0], borders(T)=0..T; streaming borders == batch select_positions within each mode; streamed LogDet gains == the batch values at the batch path’s own f32 round-trip	2e-7 rel (the f32 output of marginal_gains_exact is the floor)

Falsify (the gate can fail, proven): the missing-½ and ε²→ε⁴ mutants go red against the direct route; the static-stride chunker — the v0 this lane replaces — misses all three planted bursts. NOT asserted: cross-mode border equality — the paper’s own Table 4 prices L2 vs log-det at ~0.01 BPB of VALIDATION loss, not rank-identity at init (the same finding byteflow_rate_oracle.rs already carries; on the burst stream they disagree only on one non-burst border).

Why these two tests are the strongest available tier: the paper’s code is not public (byteflow-rate-2026-10-02.md §3, tier (b) — prospected empty); the fixture-based oracle that lane wrote already pins Rust-vs-numpy(eigen) at 7.8e-8. This round’s gates add the route the fixture oracle does not check: the T×T literal form against an LU route, computed inside the test, no fixture file to rot.

3. The integration — --byteflow, off by default

• Config (dormouse-core/src/config/schema.rs): use_byteflow: bool (default false) + the byteflow_* hyperparams, serde-defaulted to the byteflow_9m build (d_local 96, d_global 768, k_tokens 128, e_layers 2, g_layers 2, heads 4/8, w_local 256, d_ff 256/2048, bins 16, eps2 0.5, max_bytes 512); all reachable from --set/preset (the override match’s unreachable-field trap does not grow).
• Refusals, always naming the escape (ADR-0011): config::validate refuses use_byteflow with any dormouse arm on (kda/engram/mor/gr/attnres/ mhc/situ / moe-tuning / aux weights) and max_seq_len > byteflow_max_bytes (the RoPE bound, here instead of a forward assert); the trainer (dormouse_train::byteflow::check) refuses the train-side arms (--jepa-targets, --engram-ram, --graph-capture, --rand-depth, --eval-depths, --stress, --bf16, --opt ∉ {adamw, mix}).
• The loop (dormouse-train/src/byteflow.rs, dispatched FROM train_loop after the snapshot and after the device seed — the dormouse files otherwise untouched, the NaN firewall NOT touched): ByteFlowNet on the shared data seam (ByteStream::train_and_eval keeps the no-leak rule), the honest gathered CE, loss masked by the same mask_nonfinite plus a ByteFlowNet gradient sanitizer (missing grad for a visited param = loud error), AdamW (the mix router walks DormouseModel’s tree; the step-0 arm line says what ran — params= patcher= k= bins= eps2= opt= quant=unused as COUNTED), WSD lr, rewind-before-eval with the scored byte count printed on the eval line (§2.6’s window rule).
• Checkpoints: the trainer’s container pattern with its own magic (DMBF…) — model + optim records, atomic tmp+rename, .prev hard-link rotation, full resume (record round-trip + skip_bytes fast-forward); a foreign container with the same <name> is a loud magic panic, never a silent start-over.
• Preset: configs/byteflow.toml (the arm on, every dormouse arm explicitly off so the validators stay quiet).

4. Tests (all CPU, GPU never touched — the card is the graph-stage lane’s)

suite	result
burn-byteflow crate (vendor workspace)	27 passed (16 net + 7 rate-oracle + 2 dynamic-patcher + 1 field-oracle + 1 doc-test), clippy clean on the new code
dormouse-train lib	58 passed — the dormouse path’s own tests are dispatch-untouched
dormouse-train/tests/byteflow_smoke.rs	2 passed: one_step_trains_and_saves_through_the_real_entry (1 step, batch 2 s16, flex; ckpt + ce line + config snapshot) and a_preset_arm_conflict_is_refused_loudly
dormouse-core lib	89 passed (schema/validate/load — snapshots carry the new fields)

5. QUEUED — the GPU A/B (the release gate, blocked on the card)

The 200-step comparison the byteflow2 findings doc’s §6 spec’d:

• control: the byte-level dormouse recipe (--preset small, aux 0, batch 8, seq 512, seed 1, fp32) — rebuilt on a card where the attention arm actually trains (the post-8fa5d4c precondition in AGENTS §3.4; read fused kda= and engram= on the control’s eval lines before believing anything).
• arm: --preset byteflow --byteflow --batch 8 --seq-len 512 --seed 1 (AdamW per the arm line), same data window, same eval-batches → the SAME scored window (§2.6: the window is eval_batches × batch × seq_len — one batch size for all arms).
• metric: best held-out BPB (the log’s eval lines; the scored window is named in the comparison). Smoke-first: 20 steps both arms, NaN and speed, per AB-PROTOCOL’s ladder.
• blocked because ONE heavy thing at a time — the card is running the graph-stage lane; launch only when it is free, under the build lock for the binary build and systemd-run --user --scope -p MemoryMax=40G for the run.

6. Honest gaps, carried

1. The LogDet chunker path under AUTODIFF is compile-verified only. The crate’s own LogDet test runs on a bare (non-autodiff) device; under the trainer’s autodiff device marginal_gains_exact host-syncs an autodiff tensor (into_data) — the smoke runs the default L2 path, so the exact route under backward is UNTESTED. It is the analysis path; the L2 default is what the paper streams.
2. The arm line’s quant=unused is a COUNTED statement, not a refusal — a byteflow preset that carries --quant fp8 runs fp32 and says so. Same for use_tsct, which no byteflow reader reads.
3. rope_base 500_000 remains the TODO(бумага) carry from the crate (the paper states only “Multi-level” positional handling).
4. The VOCAB = 256 vs the paper’s 258 divergence stands as the byteflow2 findings doc left it (owner call, inert while RETAIN_TOKEN_IDS is empty).