Research — Mission-Type DRG Edges + Graph Sharding
Design resolved by three squad passes: pre-spec (architect-alphonso), post-spec (reviewer-renata + paula-patterns) on the edges, and — after graph sharding (#2680) was bundled in — a post-plan investigation squad (architect-alphonso + paula-patterns + reviewer-renata) on the two-phase mission. Grounded in docs/notes/mission-type-completion-arc-research.md §2 and the #883 brief / mission-type-resolution ADR. No open [NEEDS CLARIFICATION].
D-0 — Why sharding is bundled here, and why it runs SECOND (durable rationale, for posterity)
Operator's governing rationale (verbatim): "we need to do it anyway, so I prefer to keep it close to the change it risks impacting. that gives us the clearest warning/errors." The DRG graph must be sharded regardless (the monolithic 129 KB generated graph.yaml is a merge-conflict magnet — 47 commits have touched it, ~20 live branches carry a diff — and the mission-type arc's future slices grow it further). Rather than defer sharding to a disconnected later mission where the impact link is lost, it rides next to the mission-type edge change it most affects, so any breakage surfaces adjacent to its cause.
Why edges run first, not sharding (the post-plan squad's correction). The fold was initially sequenced sharding-first as a "tidy-first enabler" touching ~1 call-site. The post-plan squad refuted that against the code and the order was reversed:
| Finding | Evidence | Consequence |
|---|---|---|
| Blast radius ≈ 22 sites, not 1 | ~6 src readers (2 hardcode .../graph.yaml: pack_validator.py:513, calibration/walker.py:437; + charter_runtime/lint/_drg.py, agent_profiles/repository.py) + ~16 test modules; no canonical seam | Sharding is a migration, not a footnote → own phase + a seam (DD-6) |
| Sharding-first breaks the mission's own gate | _count_orphans (test_doctrine_regenerate_graph.py:47) reads the monolith directly | Edges must clear the 18→10 gate against the monolith FIRST; sharding preserves it |
3 readers swallow DRGLoadError | repository.py / pack_validator.py / _drg.py degrade to empty | Naive delete = green tests, silently-lost lineage/lint/validation → output-level proofs (DD-10) |
| Plan contradicted itself on fragment location | FR-007 src/doctrine/ vs plan block src/doctrine/drg/; the 6 safe consumers glob src/doctrine/ | Proof it wasn't code-traced → location pinned (DD-7) |
| "Localizes calibration diff" refuted | calibrate_surfaces is a global reweight | Benefit claim dropped; C-008 diff-width is unavoidable, not a bug |
Resolution: keep bundled (proximity rationale), but Phase 1 = edges (#2677) against the monolith (small, verified, clears the red gate), Phase 2 = sharding (#2680) as a behavior-preserving migration gated by a canonical seam + partition-totality + a merge-order equality contract + the three silent-degrade output proofs. Full decision chain: traces/design-decisions.md DD-0, DD-6..DD-10. The edge design itself (D-1..D-6 below) survived the squad unchanged — 21 edges, residual 10, requires, _KIND_MAP.
D-1 — One edge class resolves all 8 orphans
mission_type:X → action:X/<step> (requires), sourced from action_sequence. The 4 mission_type: nodes gain outbound edges; the 4 action:plan/ nodes (orphaned because plan's action indices are all-empty → no outbound scope edge) gain inbound edges. Non-plan action nodes already carry scope edges (non-orphan), so the class touches only the 8. Count 18→10 (≤ ceiling 14). Verified against the live graph.yaml by both post-spec agents. Total edges: 21 (software-dev 5 + documentation 7 + research 5 + plan 4).
D-2 — Relation is requires (not instantiates)
requires = composition-of-mandatory-ordered-parts; it is the relation the charter cascade traverses, so wiring it makes the currently-no-op mission-type cascade meaningful (Scenario 2 / SC-004). Grounded, not improvised. Reviewer may revisit with rationale.
D-3 — Dangling / cycle / duplicate safety
action_sequence ⊆ actions/<step>/ verified for all 4 types (0 dangling). mission_type → action cannot cycle (action nodes never point back to mission_type; only scope outbound). Dedup via existing _ensure/ _add_edge. assert_valid passes.
D-4 — _KIND_MAP decision
_KIND_MAP (extractor.py:122-131) has no "mission_type" entry; the generate-graph backfill loop silently drops endpoints missing from it. Harmless now (mission_type nodes are pre-created in Step 4b), but add the entry (safer against future partial-node states) + retire the obsolete :778 caveat.
D-5 — Calibration diff-width
New edges enter all_edges before calibrate_surfaces; regeneration may reweight existing edges → a graph.yaml diff wider than 21 lines. Expected, not a bug (freshness satisfied by regenerate+commit; validator passes).
D-6 — Boundary (deferred future scope)
mission_type → {WP templates, assets, guards} need node populations that don't exist (template_set null for 3/4 types; zero asset nodes; no GUARD NodeKind). Deferred per #883; none touch the 8 orphans.