[AdaWorldAPI]

Summary

The probe gets dtype and shape facts from Tract, which the plan's WHEN guards check against. Until now it read them from Tract's typed model. That stage labels only the values that survive typing: 2,546 of the Jina graph's 7,620. So 169 nodes whose op was registered were refused anyway, because their shapes were UNKNOWN.

dr3::FactSources now reads both Tract stages and keeps them apart:

• inference: the inference model after analyse. It keeps every source label. A fact there can be partial, so a shape is taken only when its rank and every dimension are determined. An undetermined dimension never becomes a symbol, because two undetermined dimensions need not be equal.
• typed: as before.
• merged() prefers the inference fact and fills a missing dtype or shape from the typed one. disagreements() compares the two wherever both describe a value.
• When typing fails, the inference facts are kept and the error is recorded in typed_error. Tract refuses to type a graph with an undetermined input dimension, but its inference stage still runs.

Results (Jina v5 small, inputs pinned [1,10])

	inference	typed	merged
values with a full fact, of 7,620	7,620	2,546	7,620
disagreements where both describe a value	0 (all 2,546 agree)
registry v1 coverage [WHEN-true / WHEN-not-true / unregistered]		3,751 / 169 / 3,700	4,146 / 0 / 3,474

• All 169 refusals came from missing facts.
• A further 226 nodes became registered once their dtype was known. WHEN-true counts went from 282 to 452 for Mul, 141 to 253 for Add, and 1 to 113 for Div.
• What remains unregistered is whole op groups: Unsqueeze 888, Shape 430, Gather 373, Concat 284, MatMul 253, Reshape 228.

Tests

• New: inference_stage_facts_cover_every_value_and_agree_with_typed, full_graph_facts_and_coverage (Jina, counts pinned exactly), undetermined_input_dim_gives_no_shape_and_typing_failure_is_recorded, and the unit test dr3::tests::fact_sources_disagree_and_merge. The unit test is needed because disagreements() reports 0 on every real fixture, so nothing else could show it ever fires.
• Re-pinned: unknown_shapes_refuse. The threshold T now has a Tract fact (the typed stage had merged it with Z), so the first UNKNOWN is n_greater, not n_where.

Test plan

• cargo fmt, cargo clippy --release --all-targets -D warnings
• cargo test --release --no-fail-fast -- --include-ignored: 26 passed
• 4 disable runs, each red on the named tests: merge from typed only; undetermined dim as a symbol; typed failure propagated; disagreements always empty
• examples/stages on the full model (numbers above); plan §0.1, board entry and STATUS_BOARD updated

---

Generated by Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: a7063942-78d9-4cd3-9fb1-90ce50688fa2

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-30T09:28:45.415043Z	40eaa80	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.