diff --git a/.claude/board/INTEGRATION_PLANS.md b/.claude/board/INTEGRATION_PLANS.md
index 10e86e452..f25d2512a 100644
--- a/.claude/board/INTEGRATION_PLANS.md
+++ b/.claude/board/INTEGRATION_PLANS.md
@@ -1,3 +1,6 @@
+## 2026-10-26 — D-TFLP v2: synergy revision → `.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v2.md`
+
+**Status:** PROPOSAL, docs only; revises v1 (#1785). **Board:** [entry](entries/2026-10-26-d-tflp-v2-synergy-revision.md), [STATUS_BOARD](STATUS_BOARD.md) D-TFLP. Folds GPT's six #1785 corrections; rescopes D-TFLP-4 to the source reciprocal from #1800; splits prepared-object identity into numeric cache key + semantic receipt from #1800/#1801; adds D-TFLP-5a (intake bake-off, #1802 pattern); records boundaries with DT64-5 and DR256/DQM.
 ## 2026-10-26 — index catch-up: thirteen plans merged without an index row (#1779–#1784, #1794–#1799, #1802)
 
 - D-NAM leaves (EXO-152): `d-nam-1-native-method-crosswalk-spec-v1` (#1779), `d-nam-2-onnx-tflite-leaf-dto-v1` (#1780), `d-nam-3-mlir-sql-relalg-normalization-v1` (#1781), `d-nam-4-quack-soa-dr-native-recipes-v1` (#1782), `d-nam-5-reverse-dto-certification-gates-v1` (#1783), `d-nam-palette-native-lut-capstone-v1` (#1784). Capstone `d-nam-native-algebra-dto-capstone-v1` (#1778) is already indexed.
diff --git a/.claude/board/STATUS_BOARD.md b/.claude/board/STATUS_BOARD.md
index f093e3f9e..e18cebacd 100644
--- a/.claude/board/STATUS_BOARD.md
+++ b/.claude/board/STATUS_BOARD.md
@@ -21,16 +21,17 @@
 
 ## D-TFLP — TFLite models on the native polyfill: int8 codes as Palette256 (2026-10-25)
 
-**Status: PROPOSAL, DOCS ONLY.** Plan: [d-tflp-tflite-palette-polyfill-capstone-v1.md](../plans/d-tflp-tflite-palette-polyfill-capstone-v1.md). Ledger: [2026-10-25-d-tflp-0-tflite-palette-polyfill.md](entries/2026-10-25-d-tflp-0-tflite-palette-polyfill.md). Specializes D-NAM-2 (#1780) and D-NAM-PZL (#1784); no executor, tensor IR or `Rel` variant.
+**Status: PROPOSAL, DOCS ONLY.** Plan: [v2](../plans/d-tflp-tflite-palette-polyfill-capstone-v2.md) (revises [v1](../plans/d-tflp-tflite-palette-polyfill-capstone-v1.md)). Ledger: [2026-10-25-d-tflp-0-tflite-palette-polyfill.md](entries/2026-10-25-d-tflp-0-tflite-palette-polyfill.md). Specializes D-NAM-2 (#1780) and D-NAM-PZL (#1784); no executor, tensor IR or `Rel` variant.
 
 | D-id | scope | status | gate / falsifier |
 |---|---|---|---|
-| **D-TFLP-0** | capstone, asset map, board rows | Proposed, docs only | no code or ABI change |
+| **D-TFLP-0** | capstone, asset map, board rows | Shipped (#1785); v2 revision docs only | no code or ABI change |
 | **D-TFLP-1** | ndarray: gemmlowp `tanh`/`logistic`/`exp_on_negative_values` + 256-entry Tanh/Logistic tables | Queued (NOT RUN) | bit-identical to TFLM `tanh.h`/`logistic.h` over all 256 codes, ≥ 5 parameter sets; disable run red |
-| **D-TFLP-2** | ndarray: 65,536-entry ADD/MUL tables | Queued (NOT RUN) | bit-identical to `add.h`/`mul.h` over all pairs; disable run red |
-| **D-TFLP-3** | ndarray: FC/conv via bias fold + existing GEMM + `requantize_i32_to_i8` | Queued (NOT RUN) | bit-identical to `fully_connected.h`/`conv.h`; accumulator range for the model's k |
-| **D-TFLP-4** | ndarray: softmax diff-keyed exp table + row reduction | Queued (NOT RUN) | bit-identical to `reference/softmax.h` int8 path |
-| **D-TFLP-5** | lance-graph-tflite facade (ALIAS/INLINE/TABLE/REFUSE), cold importer, bind once | Queued (NOT RUN) | ndarray-only native dep; identity-keyed tables, never shape-keyed |
+| **D-TFLP-2** | ndarray: ADD/MUL as 65,536-entry table **or** fixed-point SIMD, faster kept per target | Queued (NOT RUN) | bit-identical to `add.h`/`mul.h` over all pairs for both arms; ns per arm; SRAM-budget refusal |
+| **D-TFLP-3** | ndarray: FC/conv via bias fold + existing GEMM + `requantize_i32_to_i8` | Queued (NOT RUN) | bit-identical to `fully_connected.h`/`conv.h` incl. padding = input zero point, groups/dilation/stride/edges; `weights_offset≠0` refuses; accumulator range for the model's k |
+| **D-TFLP-4** | ndarray: softmax prepared (A, B) table + per-row algorithm with the **source** reciprocal; exact reciprocal only behind a PROVED receipt | Queued (NOT RUN) | #1800 replay fixtures pass; `--reciprocal exact` fails all 556 witnesses |
+| **D-TFLP-5a** | TFLite intake bake-off (FlatBuffer reader options), #1802 pattern | Queued (NOT RUN) | coverage report + adversarial fixtures; no native crate depends on the importer |
+| **D-TFLP-5** | lance-graph-tflite facade (ALIAS/INLINE/TABLE/REFUSE), cold importer, bind once | Queued (NOT RUN) | ndarray-only native dep; prepared objects carry numeric cache key **and** semantic receipt; same table under a different receipt refuses |
 | **D-TFLP-6** | op census on pinned real int8 models | Queued (NOT RUN) | per-model ops by verdict |
 | **D-TFLP-7** | one whole model end to end | Queued (NOT RUN) | output bit-identical to TFLM; load µs, memory, hot-path ns reported |
 | **D-TFLP-8** | domain-B Palette256:palette256 cosine tables (PZL-3) | Queued (NOT RUN) | PZL gates, EXO-132 k-hop closure |
diff --git a/.claude/board/entries/2026-10-26-d-tflp-v2-synergy-revision.md b/.claude/board/entries/2026-10-26-d-tflp-v2-synergy-revision.md
new file mode 100644
index 000000000..eb44d4658
--- /dev/null
+++ b/.claude/board/entries/2026-10-26-d-tflp-v2-synergy-revision.md
@@ -0,0 +1,16 @@
+# 2026-10-26 — D-TFLP-V2: synergy revision after #1786–#1803
+
+**Status:** PROPOSAL, docs only. Plan: `.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v2.md`.
+
+Reconciles v1 (#1785) with GPT's six review corrections, the 6.5 lab evidence (#1800 softmax
+certification, #1801 8-bit admission, #1803 compaction), D-ONNX-INTAKE-0 (#1802), DT64 (#1794–#1799),
+DR256/DQM (#1787–#1792) and ndarray #376/#377.
+
+Changes to stage rows: D-TFLP-0 shipped; D-TFLP-2 measures table and fixed-point arms; D-TFLP-3 adds
+padding = input zero point and `weights_offset` refusal; D-TFLP-4 ports the source reciprocal (13/14
+known and 8/8 held-out calibrations require it, lab), exact reciprocal only behind a PROVED receipt;
+D-TFLP-5a added; D-TFLP-5 carries cache key + receipt.
+
+OPEN: generation identity owner (EXO-152); binding store (D-NFR S5); `Receipt` contract placement;
+accumulator wrapping; full-runtime parity; target models; Mish subgraph validation; five unread TFLM
+kernels.
diff --git a/.claude/board/entries/README.md b/.claude/board/entries/README.md
index c33c9d584..8a3e991ce 100644
--- a/.claude/board/entries/README.md
+++ b/.claude/board/entries/README.md
@@ -25,7 +25,7 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-485 entries, 2026-08-06 .. 2026-10-26.
+486 entries, 2026-08-06 .. 2026-10-26.
 
 | date | entry id | finding | file |
 |---|---|---|---|
@@ -36,6 +36,7 @@ exactly one of them, never both.
 | 2026-10-26 | `D-DR256-OD` |  | [2026-10-26-dr256-on-demand-l1.md](2026-10-26-dr256-on-demand-l1.md) |
 | 2026-10-26 | `D-DR256` |  | [2026-10-26-dr256-global-union-lab.md](2026-10-26-dr256-global-union-lab.md) |
 | 2026-10-26 | `D-DQM` | FAIL | [2026-10-26-dqm-p1-chunk-touched.md](2026-10-26-dqm-p1-chunk-touched.md) |
+| 2026-10-26 | `D-TFLP-V2` |  | [2026-10-26-d-tflp-v2-synergy-revision.md](2026-10-26-d-tflp-v2-synergy-revision.md) |
 | 2026-10-26 | `D-TFLP` |  | [2026-10-26-d-tflp-65-landing.md](2026-10-26-d-tflp-65-landing.md) |
 | 2026-10-26 | `coresearch-lance13-masked-read-dr64` |  | [2026-10-26-coresearch-lance13-masked-read-dr64.md](2026-10-26-coresearch-lance13-masked-read-dr64.md) |
 | 2026-10-25 | `D-DRF-W8` |  | [2026-10-25-drf-w8-budget.md](2026-10-25-drf-w8-budget.md) |
diff --git a/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v1.md b/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v1.md
index 0779fee3c..d3ebc06cf 100644
--- a/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v1.md
+++ b/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v1.md
@@ -1,5 +1,7 @@
 # D-TFLP — TFLite models on the native polyfill: int8 codes as Palette256, exhaustive tables (capstone v1)
 
+> ⊘ **Revised 2026-10-26 by [v2](d-tflp-tflite-palette-polyfill-capstone-v2.md).** §1 items 1–3 (exactness), §3.1 Mish, §3.2, §3.3 softmax, §3.4 padding, §4 identity, §7 D-TFLP-2/-3/-4 and §9 are corrected there. §1–§2 and §5 stand.
+
 **Date:** 2026-10-25. **Status:** PROPOSAL / DOCUMENTATION ONLY. No code, no ABI, no
 crate in this PR. Every stage below is NOT RUN and is approved separately.
 **Parent arcs:** D-NAM ([EXO-152](https://linear.app/exo-red/issue/EXO-152), #1778–#1783),
diff --git a/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v2.md b/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v2.md
new file mode 100644
index 000000000..37780aed0
--- /dev/null
+++ b/.claude/plans/d-tflp-tflite-palette-polyfill-capstone-v2.md
@@ -0,0 +1,182 @@
+# D-TFLP — TFLite models on the native polyfill (capstone v2: synergy revision)
+
+**Date:** 2026-10-26. **Status:** PROPOSAL / docs only. No crate, kernel or ABI change.
+**Supersedes:** v1 (`d-tflp-tflite-palette-polyfill-capstone-v1.md`, #1785) where this file says so.
+v1 §1–§2 (the key fact, the asset map) and §5 (two code domains) stand unchanged and are not
+restated. Every other v1 section is listed below as KEPT, MERGED, CLOSED, EXCLUDED or EXPANDED.
+**Inputs reconciled:** GPT review of #1785 (six corrections); #1786–#1803 on `main` `13e370261`;
+ndarray `master` `af4910a` (#376, #377).
+
+## 0. What changed since v1, in one table
+
+| source | what it adds to D-TFLP | effect |
+|---|---|---|
+| GPT review of #1785 | 6 corrections (exactness, fusion, conv padding, softmax algorithm, table cost, kernel/build parity) | folded into §2 below |
+| #1800 softmax certification (lab) | certified preparation gate; TFLM-free replay crate; 14+8 calibrations decided | D-TFLP-4 rescoped (§3.3) |
+| #1801 8-bit admission (lab) | `Receipt` / `Expect` / `bind`, 6×6 codec matrix, only the diagonal binds, 17 tests | prepared-object identity (§4) |
+| #1803 | evidence compaction; run record regenerable, not kept | none on decisions |
+| #1802 D-ONNX-INTAKE-0 | cold-frontend bake-off pattern, adversarial fixture list, ORT as oracle | same shape for the TFLite importer (§4, D-TFLP-5a) |
+| #1794–#1799 DT64 | DT64-5 maps "TFLite int8 nonlinear → prepared code LUT, source-compatible"; DT64-5D = one ONNX/TFLite leaf matching D-NAM | shared leaf; keeps predicate bits and Palette256 apart (§6) |
+| #1787–#1792 DR256 / DQM | DuckDB machine-room reuse plan cites v1 open point 5 (pair table vs fixed-point SIMD) and keeps "Softmax is not a unary table" | confirms §3.2, §3.3 |
+| ndarray #376 / #377 | `masked_sum_i32` (dense branch-free / sparse walk), `mask_scatter_or_u16` | candidates for masked SUM/MEAN and u16 index maps (§3.5) |
+
+## 1. Exactness claim (corrects v1 §1 items 1–3)
+
+v1 said a table is "bit-exact by construction". Corrected: **a table is exact relative to its
+generator.** It is bit-exact once its generator has been exhaustively compared against an
+independently compiled TFLM reference, at the pinned revision, for that operator, parameter set and
+build mode. A Rust port is not self-certifying. Every TABLE verdict therefore names (operator,
+parameter set, TFLM revision, compiler/build mode) and the comparison that admitted it.
+
+Chains stay closed only where TFLite requantizes to int8 between ops; a `DEQUANTIZE` is a float
+boundary, not a closed table.
+
+## 2. The six GPT corrections, where each lands
+
+| # | correction | lands in |
+|---|---|---|
+| G1 | table exact only after exhaustive comparison to pinned TFLM, per op + params + build mode | §1; every D-TFLP gate |
+| G2 | fusion needs matching intermediate quantization, wiring, no side effects, observable outputs; Mish fuses only if the imported subgraph is deterministic, single-input, elementwise | §3.1 Mish |
+| G3 | conv padding: TFLM skips out-of-image positions (`conv.h:87-92,199-204`); dense GEMM + bias fold needs padded cells equal to the input zero point; test groups, dilation, stride, edges; refuse per-tensor `weights_offset ≠ 0` | §3.4, D-TFLP-3 gate |
+| G4 | softmax is a table plus a per-row algorithm (`softmax.h:75-133`: `diff_min`, `kScaledDiffIntegerBits=5`, `kAccumulationIntegerBits=12`, `GetReciprocal`, `RoundingDivideByPOT`) | §3.3 |
+| G5 | TABLE is a binding-time strategy; a 64 KiB pair table may lose to SIMD arithmetic or exceed MCU SRAM; measure | §3.2, §7 |
+| G6 | parity is a pinned kernel/build claim; `simd_fixedpoint` double-rounding is the default, its shift/range contracts go into the oracle matrix | §5 |
+
+## 3. Operator classes (v1 §3, revised)
+
+Verdict vocabulary KEPT: ALIAS / INLINE / TABLE / NATIVE-TODO / REFUSE.
+
+### 3.1 Unary → TABLE (KEPT, one correction)
+
+`TANH`, `LOGISTIC` read. Still candidates (kernels not read): `HARD_SWISH`, `ELU`, the `RELU`
+family, `QUANTIZE`/`DEQUANTIZE`. **Mish (corrected per G2):** a converted Mish is fused into one
+256-entry table only after the importer shows the actual subgraph is deterministic, single-input and
+elementwise with matching intermediate quantization; otherwise each builtin stays its own table, and
+a fused custom op stays REFUSE.
+
+### 3.2 Pair → TABLE or fixed-point SIMD (MERGED with v1 open point 5)
+
+`ADD`, `MUL` read. The choice between a 65,536-entry table and direct `simd_fixedpoint` arithmetic is
+made at binding time per target and measured (G5). Both are admissible only after the G1 comparison.
+On MCU targets the 64 KiB table is refused when it does not fit the stated SRAM budget. The DQM
+reuse plan cites this same open point; D-TFLP-2 answers it for both.
+
+### 3.3 Softmax (REWRITTEN from #1800 evidence and G4)
+
+v1 said "exp table + one reduction". Corrected:
+
+- **Prepared once:** the `(admitted, A = exp_on_negative_values, B = Rescale<12>)` table over
+  `d = −255..0`, from `(mult, shift)` of `PreprocessSoftmaxScaling(β, scale)`.
+- **Per row:** `diff_min` skip, row sum `S` (Q12.19), reciprocal, `RoundingDivideByPOT`, output
+  offset and clamp — the per-row algorithm, not a table.
+- **Reciprocal:** default is the **source** reciprocal (`GetReciprocal`). The exact reciprocal is
+  admitted only for a calibration the certified gate decided `PROVED_OUTPUT_EQUIVALENT`. Measured on
+  main (lab): of 14 known calibrations, 1 proved (calibration 9) and 13 require the source reciprocal;
+  all 8 held-out require it. The saving is about 26 ns per row on the reciprocal alone and is not
+  resolvable for the whole recipe at n ≥ 256. **D-TFLP-4 therefore ports the source reciprocal; the
+  exact path is an optional, receipt-gated specialization, not the plan.**
+- **Domain:** int8 input, output scale 1/256, zero point −128, `S ∈ [2^19, 2^28)`; outside it, REFUSE.
+- **Oracle:** the #1800 replay fixtures (529 full / 27 reduced witnesses, 7 equivalence rows, 10
+  unreachable) become D-TFLP-4's CI oracle, so CI needs no TFLM build.
+
+### 3.4 FC / conv (KEPT + G3)
+
+Bias fold KEPT. Added (G3): padded positions must equal the input zero point before a dense GEMM is
+equivalent; else the binding uses TFLM's skip semantics. Gate adds groups, dilation, stride, image
+edges, and REFUSE for per-tensor `weights_offset ≠ 0`. Accumulator range (v1 open point 1) is MERGED
+with #1801 OPEN item "accumulator wrapping": biased `VPDPBUSD`/`TDPBUSD` wrap by design; the gate
+proves the model's `k` stays in range, and the `wrapping_*` proposal stays on its research branch.
+
+### 3.5 Reductions and data movement (EXPANDED)
+
+Masked `SUM`/`MEAN` over admitted rows can use ndarray `masked_sum_i32` (#376); `u16` index maps for
+small-index `GATHER` can use `mask_scatter_or_u16` (#377). Both are **candidates**, exact only after
+the G1 comparison; neither is asserted to match TFLM rounding.
+
+### 3.6 Refused (KEPT)
+
+Custom ops, Flex ops, control flow, dynamic shapes, int16 (16×8), sparse weights. EXPANDED with the
+#1802 adversarial list where it applies to TFLite: unknown op/version, unknown options, mismatched
+shapes, truncated buffers, int8/u8 signed zero, bool tensors.
+
+## 4. Identity of prepared objects (REWRITTEN from #1800 + #1801)
+
+v1 said "keyed by model digest, subgraph, index and parameter set". Corrected into two keys:
+
+- **Numerical cache key:** what the table is a function of (softmax: `(mult, shift)`; unary: the
+  requant parameters). Equal keys mean identical bytes, and a table may be shared. Two calibrations
+  can prepare the same table (measured: known calibrations 1 and 2).
+- **Semantic receipt:** what the result is issued for — codec signature, scale/zero point, operator,
+  source revision and digest, codebook generation. Never derived from the cache key. #1801's
+  `Receipt`/`bind` is the lab shape: of a 6×6 producer/consumer codec matrix, only the diagonal binds
+  (rounding, code space, storage and NaN-policy mismatches all refuse).
+
+A TABLE binding needs both: cache key for sharing, receipt for refusal. This replaces the v1 "never by
+shape alone" rule (which still holds, Burn precedent).
+
+**Not decided here:** who issues `codebook_generation` (EXO-152 / PZL-4); where a bound receipt lives
+(D-NFR S5 binding store, MISSING); whether `Receipt` enters a contract type.
+
+## 5. Reference truth (v1 §6, EXPANDED by G6)
+
+TFLM reference kernels at fork `8f6596d`, compiled offline. Every parity claim names kernel, revision
+and build mode. ndarray `simd_fixedpoint` double rounding is the default arm; its shift and range
+contracts are copied into the oracle matrix. Full-runtime parity stays UNVERIFIED (v1 open point 2,
+KEPT).
+
+## 6. Boundaries with neighbouring plans (NEW)
+
+- **D-ONNX-INTAKE-0 (#1802):** same cold-frontend rules apply to the TFLite importer: no native crate
+  depends on it; unknown nodes refuse; importer IDs are not stable source identities; no automatic
+  conversion to Palette256.
+- **DT64-5 (#1798):** TFLite int8 nonlinear is a "prepared code LUT, source-compatible" method leaf.
+  Execution metadata (code space, −128 policy, rounding) never enters the 64 DT64 predicate slots;
+  Palette256 (256 codes) and the predicate catalogue (64) are distinct. DT64-5D ("one ONNX method leaf
+  matching D-NAM") and D-TFLP-5 share the leaf contract.
+- **DR256 / DQM (#1787–#1792):** RelAlg is not extended with Tanh, MatMul or Conv; nothing in the hot
+  path matches a TFLite operator name. Unchanged.
+- **D-NAM-PZL (#1784):** domain B, unchanged (v1 §5).
+
+## 7. Stages (v1 §7, revised)
+
+| D-id | change vs v1 | gate |
+|---|---|---|
+| D-TFLP-0 | shipped (#1785); this v2 is its revision | docs |
+| D-TFLP-1 | KEPT; first experiment | G1 exhaustive vs TFLM `tanh.h`/`logistic.h`, ≥5 param sets, disable run red |
+| D-TFLP-2 | MERGED with open point 5: table **and** fixed-point arm, keep the faster per target | G1 over all 65,536 pairs; measured ns per arm; SRAM budget refusal |
+| D-TFLP-3 | + G3 padding/zero point, groups/dilation/stride/edges, `weights_offset` refusal, wrap range | G1 + G3 cases |
+| D-TFLP-4 | REWRITTEN: source reciprocal + per-row algorithm; exact reciprocal only behind a PROVED receipt | #1800 replay fixtures pass; `--reciprocal exact` negative control fails 556 witnesses |
+| D-TFLP-5a | NEW: TFLite intake bake-off (FlatBuffer reader options), #1802 pattern | coverage report, adversarial fixtures |
+| D-TFLP-5 | facade; prepared objects carry cache key + receipt (§4) | identity tests incl. same-table/different-receipt |
+| D-TFLP-6..8 | KEPT | as v1 |
+
+Order: 1 → 2 → 4 in parallel with 3; 5a before 5; 6 once 5a reads op lists.
+
+## 8. Costs (v1 §8, KEPT + G5)
+
+Add per pair op: table bytes vs SRAM budget, and table vs fixed-point ns per element per arm. Softmax:
+reciprocal ns source vs exact, per-row cost by `n`, gate cost per calibration (lab: median 291 ms) and
+certificate cost per source revision (lab: 5.4 s).
+
+## 9. Open points (v1 §9, revised)
+
+| v1 # | now |
+|---|---|
+| 1 accumulator range | MERGED into §3.4 with #1801 "accumulator wrapping"; still OPEN |
+| 2 full-runtime parity | OPEN, unchanged |
+| 3 target models | OPEN, unchanged |
+| 4 int16 | EXCLUDED from first scope, unchanged |
+| 5 pair table worth it | MERGED into D-TFLP-2; also cited by the DQM reuse plan |
+| — | NEW OPEN: generation identity owner (EXO-152); binding store (D-NFR S5); `Receipt` contract placement |
+| — | NEW OPEN: Mish subgraph validation from a real converted model (G2) |
+| — | NEW OPEN: reading the five unread TFLM kernels (`hard_swish`, `elu`, `squared_difference`, `maximum_minimum`, `comparisons`) |
+
+## 10. Closed and excluded (for the record)
+
+CLOSED (by evidence on main, lab scope): whether the exact reciprocal can replace `GetReciprocal`
+(answered per calibration; mostly no); whether a codec-space mismatch is caught at binding (yes, 6×6
+diagonal); whether a table can silently belong to another calibration (refused,
+`table_identity_mismatch`).
+
+EXCLUDED: a TFLite runtime dependency; a tensor AST; an ONNX/TFLite → RelAlg graph; automatic domain
+A → B conversion; a Fisher-Z round trip in domain A.