fix(ce-plan): close escape hatches that let the skill abandon direct invocations

[tmchow]

Summary

ce:plan had interpretive gaps that let the agent classify legitimate planning requests as "not a planning task" and abandon the workflow, even on explicit /ce:plan invocations. This closes those escape hatches and adds smarter routing when planning isn't the right tool.

What changed

Closed escape hatches (commit 1):

• Removed the "prefer ce:brainstorm first" sentence from the description frontmatter. It was meant for auto-selection routing but gave the model justification to reroute on direct invocation too.
• Added a top-of-body directive: "When directly invoked, always plan." with explicit instruction to never abandon. This is the primary fix.
• Replaced the dead-end "Do not proceed until you have a clear planning input" with actionable alternatives: ask clarifying questions or use Phase 0.4's bootstrap.
• Scoped the Phase 0.1b "everything else" exit clause (quick questions, factual lookups) to auto-selection only.
• Broadened Phase 0.4 from "No-Requirements-Doc Fallback" to "Planning Bootstrap" so it covers any underspecified input.

Smart routing after bootstrap (commit 2):

After Phase 0.4's bootstrap establishes enough context, ce:plan now detects two cases where a different workflow fits better:

• Symptom without root cause (broken behavior, cause unknown): auto-routes to ce:debug with announcement. Can't plan a fix for something you haven't diagnosed.
• Clear task ready to execute (known fix, no decisions to make): suggests ce:work as a faster alternative. User chooses.

Brainstorm suggestions and universal planning routes are preserved. The difference is they're now suggestions with continuation paths rather than gates that block the workflow.

Test plan

• bun test passes (94/94 tests, including frontmatter validation)
• bun run release:validate confirms metadata in sync
• Ran 6 simulated eval runs (3 test cases x old/new skill) to verify non-regression

---