Passes 3163–3174: proof-carrying Bayesian runtime and non-Abelian frontiers

[wilcompute]

Stack and manuscript basis

This PR is intentionally stacked on agent/pass3153-3162-adaptive-epoch-factor-engine / PR #242. It executes the five requested next fronts plus two independent outside-box fronts, after rereading the full photonic_holonet.tex source (7,332 lines) and the 38-page holonet_machine_blueprint.tex PDF.

The architecture follows those manuscripts rather than replacing them with a generic controller abstraction:

• carrier/state/operator duality and address/route duality remain explicit;
• selector clock, mirror bus, address sheets and ternary temporal extension remain distinct layers;
• support/readout is not conflated with phase/execution;
• every promoted number retains its proved/measured/derived/modelled/open type;
• M36 injection remains an external fail-closed handshake until a candidate is observed and certified;
• demonstrator, fault-tolerant machine and laboratory claims remain separate.

Transport bootstrap

The complete locally validated source packet contains 32 readable files. To avoid connector-sized partial commits, it is carried initially as nine Base85/zlib chunks plus a 31-line deterministic materializer. The bundle was reconstructed locally and byte-compared against the intended source tree: 32/32 exact matches, zero omissions, zero drift.

The bootstrap workflow must:

1. require exactly nine chunks and 32 embedded files;
2. materialize the readable source tree;
3. regenerate all fast exact/model certificates;
4. pass six focused regressions;
5. push the readable files to this branch once.

Only after that push do the focused RTL/PDF lane and the two exhaustive sharded lanes become the operative evidence surface.

1. All 194 larger universal ISAs

Pass 3163 installs a 32-shard exact full-affine BFS engine for every one of the 194 universal five- and six-opcode subsets. Each record must reach all 4,199,040 affine elements and retain its full growth series, diameter and mean distance. Aggregation refuses partial shard sets.

Pass 3164 adds a three-way physical promotion contract:

• current4: mean 14.1755851337, collision probability 45/324;
• low4: mean 15.2163239693, collision probability 36/324;
• fast6: mean 13.7293695702, collision probability 63/486.

Exact consequence in the frozen path-plus-collision model: fast6 strictly dominates current4 for every nonnegative collision price. Fast6 beats low4 below 16.7016424928; low4 wins above it. Current4 remains the fail-closed hardware fallback until decoder area, timing and calibration are observed.

No global larger-ISA distance optimum is claimed before all 194 records aggregate.

2. Complete 256-shard rank-three M36 census

Passes 3165–3166 bridge the merged Pass 3125 engine into:

• 256 logical shards covering the stated 50,868,675 isotropic rank-three subspaces;
• 32 workflow buckets of eight shards;
• normalized candidate records with engine/source hashes;
• aggregation that refuses missing or duplicated shard indices;
• independent Pass 3134 projector certification;
• Pauli spectrum, fixed-frame Wigner witness, product-stabilizer lower bound, logical frame and first/second-order accepted-error slopes for accepted outputs.

Zero accepted candidates becomes a no-go statement only if all 256 normalized shards and the independent certifier complete. Before that, no-input and partial zero-hit states remain non-results.

3. Phase-coded edit-proof epoch

Passes 3167–3168 replace the single delimiter with twelve phase-coded length-five markers, one per payload phase, using twelve of the fifteen payload-unused symbols.

Exact exhaustive result:

• marker length: 5, optimal for two-edit correction because d >= 2t+1 = 5;
• marker-marker distance: 5;
• marker-to-payload-window distance: at least 5;
• corrected edits: any two substitutions, insertions or deletions;
• distinct received traces per phase: 24,845;
• total traces exhausted: 298,140;
• clean payload symbols needed after marker: 0.

The phase is carried by the corrected marker itself. Continuing edits no longer require a separate clean two-symbol reacquisition assumption.

4. Complete streamed Bayesian decision datapath

Passes 3169–3170 compile the exact sparse factor representation back into the complete 48,826-hypothesis posterior without dense posterior RAM:

1. 528-cycle factor load;
2. one 48,826-hypothesis max-log pass;
3. one 48,826-hypothesis normalized-weight/action-bin pass;
4. 184 action-bin reductions;
5. 23 action comparisons.

Total contract: 98,387 accepted cycles per decision.

For Q12 factors with a stated exp-LUT model:

• maximum posterior error: 4.086462578367688e-7;
• maximum action-score error: 4.439082887597312e-5 bits;
• next-action agreement: 12/12 deterministic tests.

The corresponding 1,016.394442 decisions/s at 100 MHz is a modeled design point, not observed timing.

5. RTL and manuscript integration

The readable packet includes:

• current4/low4/fast6 affine decoder RTL and fixed placement wrappers;
• phase-coded epoch decoder;
• complete 48,826-hypothesis enumerator and 3,381-correction map;
• two-pass Bayesian stream and 23×8 action bins;
• information-aware tri-ISA scheduler;
• combined protocol testbench;
• idempotent integration into w33_paper.tex, photonic_holonet.tex, holonet_machine_blueprint.tex and docs/index.html;
• a focused Icarus/Yosys/nextpnr/three-PDF evidence workflow.

Bonkers frontier A — non-Abelian belief curvature

For D4, the derived subgroup is exactly {1,r^2}. Therefore every ordered pair of nonidentity fault labels has a complete, conjugation-invariant one-bit curvature syndrome:

kappa(a,b)=1 iff [a,b]=r^2.

Exact census:

• 49 ordered nonidentity pairs;
• 25 flat, 24 curved;
• across 69 shared measured edge pairs: 1,725 flat and 1,656 curved correction factors.

This turns the noncommutative correction layer into a literal Z2 curvature detector. It is an algebraic interaction syndrome, not a measured optical field strength.

Bonkers frontier B — ISA as a control channel

Treat the opcode-to-destination map at each frame as a deterministic communication channel and average destination entropy over uniform frames/opcodes.

Exact capacities:

• current4: 1.8495219522 bits/dispatch;
• low4: 1.9050775077 bits/dispatch;
• fast6: 2.4105697281 bits/dispatch.

Fast6 carries 0.5610477759 more absolute control bits per dispatch than current4, a 30.334745% gain. This distinguishes raw collision count from absolute and normalized control capacity and supplies a second signal for the tri-ISA scheduler.

Evidence gates

Do not merge this PR before:

1. PR Passes 3153–3162: larger ISA census, epoch delimiter, and factor engine #242 is green and merged;
2. the bootstrap publishes the readable 32-file source tree and passes 6/6 focused regressions;
3. the focused lane passes Icarus, Yosys, HX8K placement, stacked idempotence and all three PDFs;
4. observed outputs are inspected and typed correctly.

The all-194 ISA BFS and 256-shard M36 census are separate exhaustive workflows. Neither is a prerequisite for source review, but neither may be described as complete before its aggregate artifact exists.

No claim is made yet for a global larger-ISA optimum, accepted rank-three candidate existence, observed RTL behavior, FPGA area/timing/power, materialized manuscripts, successful PDF builds, laboratory likelihoods, physical edit rates, heat, coherence or optical performance.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[wilcompute]

Execution and evidence boundary:

• The complete 32-file source packet was reconstructed locally from the nine committed chunks and byte-compared against the working source tree: 32/32 exact matches, zero omissions, zero drift.
• Local exact/model regression is PASS 6/6. Frozen results include the 298,140-trace phase-coded epoch census, 12/12 Bayesian action agreement, the D4 one-bit curvature theorem, the ISA channel capacities, and the tri-ISA crossover algebra.
• Bootstrap run 30886351895, job 91918315004, is queued. The readable source tree is not yet claimed to be materialized on GitHub until that job pushes it.
• The all-194 affine BFS and 256-shard M36 census remain separate exhaustive gates. No global larger-ISA optimum and no M36 no-go/candidate-existence claim may be promoted from source-complete workflows.
• The focused workflow after bootstrap must still pass Icarus, Yosys, HX8K placement, stacked integration idempotence and all three PDFs.
• No observed RTL behavior, FPGA area/timing/power, materialized manuscript, PDF success, laboratory likelihood, physical edit rate, heat, coherence or optical performance is claimed from a queued state.