Harden plan-builder happy path and add unified agent preflight

[jamil2018]

Summary

• harden the plan-builder happy path across validation authoring, expected-red baseline classification and repair, implementation task projection, completion-event identity, and run-evidence counts
• add a truthful unified agent preflight to project_diagnostic that separates App/API identity, active MCP transport, current-task capability visibility, and target-project binding
• reduce repeated MCP payloads while preserving AST-ready locator and environment bindings
• synchronize scaffold templates, MCP contracts, generated coordinator references, and Graphify outputs
• add the prioritized 19-feature plan-builder implementation roadmap under codex/development plan/appraise-0.5/product direction

Why

The SecondWife happy-path audit completed end to end but exposed several points where apparently healthy state could be misleading or inefficient: server capabilities did not prove the current task's immutable MCP snapshot, historical baseline attempts exposed active controls, repeated scenario failures could be misclassified, implementation task state was incomplete at group approval, and response summaries repeated bulky historical evidence.

This change turns those findings into durable source fixes and records the remaining feature sequence.

Notable behavior changes

• project_diagnostic accepts observedTools, observedResources, and expectedTargetWorkspacePath
• zero-input diagnostics return needs_observation instead of claiming client readiness
• missing capability sentinels and target mismatches return bounded recovery guidance
• repeated approved failure signatures remain valid, but later approved signatures cannot match when the first signature never occurred
• Cucumber authored-step counts exclude runtime hooks and report hooks separately
• implementation summaries omit historical baseline attempt bodies when current implementation runs are available

Validation

• 5 focused root Vitest files: 46 tests passed
• full packages/appraisejs suite: 16 files, 125 tests passed
• npm run build
• npm run quality:fallow:commit
• npm run quality:react-doctor:commit — 100/100
• npm run check:harness
• npm run release:check:artifacts
• npm run release:check:packages
• npm run release:check:coordinator-reference
• scaffold template preparation and Graphify regeneration
• Prettier checks for changed canonical files and git diff --check

The production build passes with the repository's existing Turbopack broad dynamic-file-pattern warnings.