ultracode: fixed catalog of five default orchestration shapes

[justrach]

Problem

ultracode leaves the orchestration shape entirely to the model: it authors whatever workflow JSON it feels like. Results vary run to run, the fleet cannot tell which structure produced a score, and there is no house style for "how do we review" or "how do we research".

Proposal: a fixed catalog of five shapes, selected by task intent

A - Review / audit (default when there is a diff)

phase "find"       -> N reviewers, one per dimension (correctness, security, perf, tests), parallel
phase "verify"     -> skeptic per finding, prompted to REFUTE; when: "FINDING"
phase "synthesize" -> 1 task, ranked output

B - Research / understand (default for a question)

phase "sweep"      -> researchers, each searching a DIFFERENT way (by symbol, by call-graph,
                      by git history, by tests) - blind to each other
phase "synthesize" -> one map

C - Design / solve (already exists via scoreVariants)

phase "variants" -> N implementers with different system_prompt (MVP-first, risk-first, perf-first)
                    on the SAME task -> auto-judged -> DGM scoring round
phase "build"    -> synthesize from winner, graft runners-up

D - Migration / mechanical (pipeline, never phases)

pipeline { items: discovered sites, stages: [transform, verify], isolation: worktree }

E - Build a feature

phase "scope"     -> 1 researcher
phase "implement" -> implementers in worktrees, parallel by component
phase "review"    -> reviewer + skeptic on the diff

Tier assignment

One rule across every shape: frontier for judgment (scope, judge, verify, synthesize), mid for production (search, implement, transform), small for mechanical. Depends on the tier ladder and per-persona pin issues.

Scale knob

"find bugs" -> 3 finders, 1 verify. "thoroughly audit" -> 6 finders, 3-vote adversarial, synthesis phase.

Relationship to the fleet

Each shape slot (find / verify / synthesize / transform / ...) is emitted as the third component of the composite niche key from the archive-integrity issue. The catalog is the scaffold evolution runs inside; the slot is a coordinate of the space it illuminates.

Shapes themselves are deliberately not evolved yet - a shape is a program, not a prompt, and the sample rate would be far too low to learn anything. Once there is volume, shapes get their own coarser evolution keyed on eval_set_hash (already computed per task), reusing the same score stream one level up.

[justrach]

Depends on #291 + #292 for tier assignment per slot; emits the slot component consumed by #290.

[justrach]

Three defects found by running ultracode review src/shapes.zig against this very catalog are now fixed on feat/293-ultracode-shapes (d09f39f):

• HIGH — the catalog told dependent stages to use isolation:"worktree". Each spawn gets its own worktree from HEAD, so verify/review read the original file and edits were stranded; a workflow could report success while nothing reached the tree. Guidance now scopes it to parallel same-phase editors and forbids it on dependent chains, with a test pinning both the new wording and the absence of the old. Engine-level fix tracked in worktree isolation is per-spawn, so dependent workflow stages cannot see each other's edits #295.
• LOW — canonicalSlot("安全 review") returned review: firstWord treated each byte of a leading non-ASCII word as punctuation and promoted a later word to first, mis-filing fitness. A byte >= 0x80 during the leading skip now aborts the scan.
• LOW — transform was advertised as scoreable but pipeline stages never reach the scoring path. Kept in the vocabulary, gap now documented at the declaration. Tracked in pipeline stages are never scored, so the transform slot can never fill its cell #296.

Cold-cache baseline diff 355 -> 356. Stack (feat/290-slot-axis rebased on top) is green at 360/360.

[justrach]

Implemented and verified against real Codex, plus the three self-found defects fixed. In PR #297 (release/v0.0.217) — closing on merge.

[justrach]

Shipped in v0.0.217: 3585446 added the fixed catalog of five orchestration shapes (src/shapes.zig), and d09f39f corrected the three defects the catalog's own first review found. Shape selection has since fed the fleet slot axis (#290), and pipeline stages are scored so shape D's transform slot fills its cell (14ce015, v0.0.226).