docs(defork): design cli_run.rs resolve/reconcile/discovery -> .dag grounding (Lane 4b / Chunk F)

[gunbai-bot]

Summary

Design-only doc for the new lane assigned via sunny-newt-884 (operator directive 2026-07-02): give cli_run.rs's resolve/reconcile/discovery engine (resolve_entry_graph, reconcile_with_typed_cache, resolve_transitively, discover_owned_data_decls, resolved_graph_from_sources, whole_tree_resolved_ctx) a .dag authority, folding CI-repair Lane 4b (per-entry reconcile memo) into the seed-shrink burn-down instead of hand-patching Rust we're deleting.

No implementation in this PR. Scopes 3 concrete redundancies to dissolve (not relocate) and proposes a 3-phase authoring order:

1. Module dependency graph — repoint the regex import-scan onto decl_facts (Whole-tree EXTRACT for model-grounding lens: concept_decl_facts(roots), name index, layer exclusion, worklist schema #5966, already merged) + std.graph's existing CallGraph/adjacency BFS, instead of a bespoke text-scan and hand-rolled closure walk.
2. Cache-key policy — collapse 3 independently-keyed hand-rolled HashMap caches into one cache_interface.dag Realization instance (already modeled, uninhabited).
3. Whole-tree batch grouping — de-fork discover_owned_data_decls's closure-grouping against v2.lens.affected_set/ci_floor_plan.dag batching instead of a third bespoke algorithm.

Fenced away from src/v1/03_resolve.dag/04_infer.dag (resolver-B, PR #6155, open) per the standing infer-stage fence.

Test plan

• Docs-only change; no code touched. cargo fmt/clippy/tests unaffected.
• Reported to sunny-newt-884 (parent) for design sign-off before any implementation/cutover PR, per the brief's "design-first, escalate before cutover" instruction.

[gunbai-bot]

Both findings confirmed and fixed in 8c137a5:

1. Wired the census back-link: seed-shrink-census.md Chunk F row now points to this doc, instead of an unwired orphan claim.
2. Fixed the cache-inhabitant citation: content_addressed.dag only defines CasAddressingConfig, no CacheInterfaceCatalogFacts row — wrong cite. Replaced with extdeps/realization/parse_table_memo.dag, which is an existing locality: InProcess inhabitant — actually a stronger citation than the one I had, since it's the exact shape Phase 2's in-process cache rows should take.

— sent from still-crab-700

[gunbai-bot]

Both findings confirmed and fixed in beafec1:

1. "Six functions" was stale wording left over from the original brief — §0's own table already named 10 distinct fns. Reworded to state the actual count (15 fns / ~450 LOC: the 6 brief headline fns + their plumbing) instead of repeating the wrong number.
2. Confirmed build_multi_entry_index (cli_run.rs:673), resolve_entry_with_index (:682), resolve_entry_with_index_for_discovery_corpus (:695), resolve_entry_graph_with_index (:712), and load_sources_for_entry_with_index (:570) are real load-bearing plumbing in the same cluster, correctly not listed as headline fns but wrongly excludable under the old §4 wording. Added them to §0's table and reworded §4's fence to cover the resolve/reconcile/discovery cluster (not a literal function count), so Phase 2 isn't self-blocked by its own coordination note.

— sent from still-crab-700

[gunbai-bot]

All three findings confirmed and fixed in a155da0 -- this one was a real modeling error, not just wording:

1. Confirmed decl_facts/concept_decl_facts (src/v2/std/concept_index.dag) carries {qualified_name, name, kind, node, rel_path} -- no import edges. Wrong authority cited.
2. Confirmed the real structural surface is import_resolution_facts/module_declaration_facts (cli_run.rs ~7012-7068), already projected in src/v2/lens/module_graph.dag as ImportResolutionFact/ModuleDeclarationFact.
3. Confirmed src/v2/lens/module_graph.dag already has a working import_closure/import_closure_live fixpoint over that exact shape -- Phase 1's proposed new compile_module_graph.dag + std.graph-based closure would have been a third module-closure concept, the precise §2 violation this whole lane exists to avoid.

Rewrote Phase 1: it's now a pure repoint of cli_run.rs's closure consumer onto the existing import_closure_live, no new authority. Also dropped the Phase 1 <-> decl_facts_project.rs (#6158) sequencing gate since it was based on the wrong authority -- Phase 1 can start independently now. Noted the extract_import_paths regex-scan (still live inside import_resolution_facts) and the FreeMonoid-fixpoint vs std.graph-adjacency duplication as real but separate, out-of-scope levers.

— sent from still-crab-700

[gunbai-bot]

Investigated -- this is not a real failure. The ci job was CANCELLED at exactly the 30-minute job-timeout mark (deliberate fail-fast cap, #6154), while rust_tests SUCCEEDED and mergeable=MERGEABLE. Per operator-confirmed policy (2026-07-02): dark-CI is the designed operating mode until the floor gate itself drops under 30min (resolver-B work in flight) -- the 30-min cap is intentionally set below the ~47min floor duration to prevent the prior 360-min runner-starvation deadlock, not a bug to fix-commit against. No code change needed; PR is merge-ready under the dark-CI criteria (2/2 distinct approvals, no REQUEST_CHANGES, mergeable=CLEAN, rust_tests green) -- flagged for the operator's manual-merge queue.

— sent from still-crab-700