cognitive-shader-driver: emit CE64 from the SPOFC record (follow-up to #1293)

[AdaWorldAPI]

Follow-up to #1293 (merged); retargeted to main.

This is the follow-up to #1293. #1293 collects each row's supporting relationships into one SPOFC record, but that record never reaches the emitted CausalEdge64 (CE64). This PR writes it into the CE64 (leg A). It also verifies what the cross-cycle learning would rest on, and records what is missing for the rest rather than inventing substitutes.

Leg A: emit CE64 from the SPOFC record

Stage [5] used to pack every edge from the hit alone. Each field changes as follows:

field	before	now
S / O	row%256, (row/4)%256 (a row id placed in palette-address space)	palette256 indices: S is the row's stored archetype (the value the cascade queries with); O is the partner's (a P64 target, or a row partner's stored archetype)
P	0	0 (no predicate palette exists)
f / c	resonance, resonance	SPOFC truth: c = m·255/(m+k) counts supporting relationships
Pearl mask	predicates & 0b111	inverse of p64_bridge::edge_to_layer_mask's causal half
topology (bits 59–60)	implicit 0 = Direct, for every edge	Direct for a P64 target, Unknown for a content-similarity partner

A bug fixed on the way. p64 plane 2 is SUPPORTS, but edge_to_layer_mask reads Pearl bit 2 as CONTRADICTS. So every supporting relation was emitted as a contradiction.

Tests. Each was checked by disabling what it protects and confirming it goes red:

test	what it pins	disable that turns it red
emitted_edge_carries_the_p64_target_and_its_evidence	S = stored archetype, O = the P64 target, evidence confidence, a SUPPORTS plane sets no Pearl bit	S from the row id; c = resonance; raw mask at the call site
content_partner_edges_do_not_claim_a_direct_topology	content partners are Unknown and carry evidence confidence	topology forced to Direct; c = resonance
causal_mask_round_trips_through_p64_layers	all 256 predicate masks round-trip through edge_to_layer_mask	raw mask inside the helper

The fixture stores archetype 2 rather than 0, because row 0's row%256 is also 0 and would pass by accident.

Effect. Emitted edges change in every dispatch. cycle_fp and top-k are unchanged, since they read the candidates, not the edges.

Verified for the cross-cycle legs (not implemented)

tests/ce64_recount_probe.rs characterises current behaviour:

• learn() pools a repeated observation as fresh evidence. Confidence rises again on the second revision with the same observation.
• The driver emits with plasticity ALL_FROZEN, so learn() can never move S/P/O on those edges.

More, all read in code:

• Nothing in run writes an edge back. persist_cycle writes only emitted_edges[0], has test-only callers, and targets an Arc<BindSpace> that run cannot mutate.
• Stage [3] computes tables.revise(...) and throws the result away.
• arm-discovery conventions: the SPOFC confidence convention matches; the frequency does not. SPOFC f is the best resonance; arm-discovery's f is cooccur/antecedent, and a dispatch has no antecedent count to compute that ratio from.
• R2IL: it has no frequency/confidence code. ogar-r2il and r2il-mask-abi-probe have zero hits. "r2il SPOFC" appears only in docs.
• The EWA sandwich (jc::ewa_sandwich) is covariance push-forward and counts no evidence. Its only epistemic use is the PROPOSED mul-ewa-trust-propagation-v1.md, which names the circularity (Σ → trust → gate → Σ). A completion it supplies is a prior, never an observation.
• Shared upstream. fusion.rs (shared_roots/inherited_roots) and nars::belief::revise_at already treat two supports from one root as one witness. The dispatch path has no carrier for that.

Missing dependencies (named in the plan, not substituted)

1. W4a owner write path. COMPONENT-MAP.md §6 has persist_cycle BLOCKED→W4a. Write-back must go through cast(on_behalf = mailbox_owner()).
2. A per-edge evidence-identity carrier. The W-slot is a witness corpus root, not a per-observation source set, and le-contract.md forbids new awareness semantics in CE64 bits. Where the stamp lives is undecided.
3. A predicate palette, for the P byte.
4. A production feed into the planes. update_planes has no production caller, and convergence.rs addresses with %64 where p64-bridge uses /4.

Plan: .claude/plans/ce64-spofc-learning-v1.md.

• Legs B–F: evidence identity, write-back, cross-cycle revision, the "Sudoku" completions probe, and support entropy.
• It also sets out the framing: bits 59–60 as the grid recording how each cell was obtained, and 8× support as 8× evidence only if the eight are independent.

Checks

• cognitive-shader-driver: all test binaries pass (115 lib tests plus the new probe); clippy -D warnings and fmt are clean.
• Board: INTEGRATION_PLANS.md updated (new entry prepended); SUPERSESSION-INDEX.md regenerated.

🤖 Generated with Claude Code

https://claude.ai/code/session_019HnekoM1EidTwQLS3oFVFm

---

Generated by Claude Code

---

Generated by Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Essentials

Run ID: 13d3e1eb-e665-40b5-97dc-bc7925f06306

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_f8072d30-9f16-4d49-aefc-a4c6a3694564)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_5c84e4c7-9e66-4669-9e99-580ea9a8f2f8)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 04315c7432

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Derive the emitted predicate from the SPOFC relation

For a SUPPORTS-only dispatch using analytical style, as in the new Ordinal(1) fixture, layers_to_causal_mask emits zero while style_ord_to_inference emits Deduction; consequently p64_bridge::edge_to_layer_mask reconstructs CAUSES rather than SUPPORTS. Other non-Pearl predicate layers are similarly preserved only when the unrelated thinking style happens to select the matching inference type, so feeding emitted edges back through the documented bridge silently moves relationships between planes. Encode relation-derived inference or emit the predicates separately instead of using the style for this field.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep confidence scoped to the emitted object

When a candidate has supports for different partners—the normal cascade returns up to four distinct targets, as the new test demonstrates—spofc.object names only the best target while this confidence counts every target in c.support. The resulting CE64 therefore asserts one specific S/P/O statement with the evidence weight of unrelated alternative objects, and downstream revision or ranking will treat those alternatives as repeated confirmation of the chosen edge. Compute confidence from supports matching the emitted statement or emit separate edges for the distinct relationships.

Useful? React with 👍 / 👎.