Slim spec-to-plan master plan.md; lock the lean shape in evals

[kreneskyp]

Why

The /spec-to-plan master plan.md had become an over-stated mirror of the bundle: the same task↔track↔FR↔TC↔status matrix was restated 3–4 times in prose (Requirements Summary, Dependency Graph, a duplicate Execution Tracks section, Remaining Work per-task Scope, Test Plan, Task File Mapping) — re-deriving data the task frontmatter already holds authoritatively (references/depends_on/verifies). On a 6-task plan that's ~160 lines, and it scales badly.

Root cause was the skill's own step-1/2/3 output templates, which stacked the redundant sections; the generation then piled on more (a standalone Execution Tracks section the template never asked for, the gate stated twice, free-form The seams prose).

What

Make plan.md an orchestration overview, not a mirror — state each fact once.

Skill templates (skills/spec-to-plan/):

• step-3: Remaining Work collapses to one-liners (no per-task Scope, no TC-id exit lists — exit criteria are behavioral); a single Task File Mapping table is the only per-task matrix; no separate Execution Tracks section; gates stated once, inline in their track; new "Mirroring frontmatter in prose" Common Mistake.
• step-1: The seams promoted to an official optional subsection; noted as the single dependency-graph representation.
• step-2: Test Plan is the single TC enumeration.
• SKILL.md: states the orchestration-overview principle up front.

Evals: EV-027 gains a fileContains shape check — a single Task File Mapping table is present and there is no separate Execution Tracks section — to lock in the slimming. Eval-matrix row (spec/evals.md) updated to match.

Verification

• New EV-027 assertion verified against real fixtures: it catches the old duplicated shape (Plan-004, which carries ## Execution Tracks → fails) and passes the lean shape. (Scenario file parses; regex semantics mirror assert.mjs.)
• eslint clean, prettier --check . clean, build present.
• Live agent-pty EV-027 not run in CI/this PR: the harness's inner agent (--permission-mode bypassPermissions) blocks on Claude Code's bypass-permissions accept interstitial in environments where it isn't pre-accepted, so it must be run where that's configured. The assertion logic itself is verified offline against real plan files (above). Skill template changes are prose-only and don't touch src/.

🤖 Generated with Claude Code