feat(pipeline): give docs and skill-file changes an area in feature-factory

Start by reading .claude/skills/feature-factory/SKILL.md and the linked decisions in #400, #437, #448, and #451. Grooming must first record which of the three proposed approaches the owner chooses; then update the skill and, for option 1, add .claude/agents/docs-builder.md. Done means the scope boundaries and pipeline behavior match the chosen approach, a docs-only run reaches stop 3, and a later real run is recorded on the issue.

Follow-up to #438 (the merged feature-factory skill). Found on the pipeline runs for #456 and #444. Not groomed: it needs one decision (below) before a builder starts. Branch from main and open the PR against main.

Problem

The pipeline knows three areas: api, web and assistant. The Scope block lists which of them change (.claude/skills/feature-factory/SKILL.md, stage 3 step 1) and stage 4 runs one builder per listed area. A change that touches only docs or Claude config files (docs/, .claude/skills, .claude/agents, .claude/rules, README files) has no area and no builder.

Both runs hit it, and both handled it by hand:

• #456 (a skill edit): "The SKILL.md edit was applied by a general-purpose agent, since no builder covers a .claude skill file." https://github.com/nevridge/TaskFlow/issues/456#issuecomment-6716756859
• #444 (a docs edit): "No feature-factory builder owns docs/; the owner chose a one-off general-purpose agent for the docs edit." https://github.com/nevridge/TaskFlow/issues/444#issuecomment-6716895666

Costs: an extra decision for the owner at stop 2 every time, a different answer possible each run, and no path boundary. The three builders each say "never touches other areas"; a general-purpose agent has no such limit, so nothing stops it editing settings.json, hooks, compose files or CI.

Constraint from #400: claude-integration-advisor deliberately does not cover repo .claude/, CLAUDE.md or the project agents, because "tooling edits [are] made by hand, not feature work" (decision IA3). This issue has to either respect that or revisit it explicitly.

Decide in grooming

1. A docs area built by a new docs-builder agent with a path allowlist, like the other builders. Most consistent, but an eighth agent after #437 consolidated to seven.
2. A docs area defined in the skill and built by a general-purpose agent, with the path allowlist and "never touches other areas" stated in the skill's prompt, as #400 does for other project needs. No new agent file. This is the recommendation.
3. No area: the pipeline refuses a docs-only diff, names the reason at preflight or after the Scope block, and tells the owner to use a plain builder session from the issue (as #448 and #451 were). Consistent with IA3, but the owner loses the stops, the validator and the notes for these changes, which they have been using.

Change (for options 1 or 2)

Edit .claude/skills/feature-factory/SKILL.md (and add .claude/agents/docs-builder.md for option 1):

1. The Scope block can list a fourth area, docs: docs/**, .claude/skills/**, .claude/agents/**, .claude/rules/**, and *.md files outside docs/ (README files, CLAUDE.md).
2. Stage 4 builds it like the others (one commit, build-docs.md), in dependency order after the code areas.
3. Boundaries: never .claude/settings*.json, hooks, .github/**, compose files, .env or any source or test file. An agent file edit keeps its frontmatter valid YAML (check with js-yaml, as #448 did).
4. Stages 5 and 6 need no change: the verifier runs no gates when no code area changed, and the validator checks the final diff against the brief. State that a docs-only run skips the code gates, so the verifier does not report "environment failure".
5. The owner is not asked at stop 2 which agent to use.

Acceptance criteria

1. The owner decision above is recorded on this issue.
2. A docs-only or skill-only issue run through /feature-factory reaches stop 3 without a stop 2 question about which agent builds it.
3. The builder for that area cannot touch the paths excluded above, and says so in its prompt.
4. The change does not alter the stops, the budgets or the three existing areas.
5. Verified on the next real docs-only or skill-only pipeline run, recorded on this issue.

Tests

No tests apply (skill and agent prompts). Say so in the PR body.

Out of scope

• The claude-integration-advisor scope (IA3 stands unless the owner says otherwise here).
• Changes to pr-triage or the user-level agents.