feat(#378): resolve intent from any spec, then proceed, assume, or park

[randomm]

Closes #378. The core of Wave 1.

The problem

/work assumed the issue told it what to build. Real backlogs don't honour that — issues are hand-written, terse, imported from elsewhere, or simply wrong. 38.3% of real GitHub issues are underspecified (SWE-bench Verified, 93 annotators, 1,699 samples) and 41.77% of multi-agent failures are specification-level (MAST, the largest single category). Underspecification is the modal case.

Two structural holes:

1. A missing verdict meant "build it." work-driver-plan.ts defaulted an absent token to NEEDS_WORK. An agent that simply forgot the marker got code written. Silence was permission.
2. Nothing asked whether the issue was TRUE. No step checked whether the named symbols exist, whether the described behaviour matches the code, or whether the work was already done. A confidently-wrong bug report got built.

What it does now

The explore step resolves intent from whatever body it is handed and grounds it:

• against the code — do these symbols exist? does the behaviour match? is it already implemented?
• against the world — does that third-party API exist with that shape?

Then it decides:

Verdict	Action
proceed	continue
proceed-with-assumptions	continue, every assumption written into the PR body
park	write no code; halt at intent-park with a machine-readable reason

Park reasons — underspecified, contradicted-by-code, already-implemented, too-large, premise-unsound — each carry their own operator action, not the generic "inspect the state file and --restart" (which for an unresolvable issue just reproduces the same park).

A park halts before plan or branch, so refusing costs one explore dispatch rather than a whole cycle ending in a bad PR. Verified: nextStep after intent-park returns handoff, so no code-writing step can run.

A missing or unreadable verdict now parks. That is the inversion.

Two design points worth arguing

Evidence overrides the resolver's own verdict. An LLM can claim proceed while recording that the issue's central claim is contradicted by the code. When it does, the contradiction wins. That contradiction is the highest-value signal this step produces, and ignoring it is precisely how a wrong bug report gets built. Tested directly.

The resolver structurally cannot build. It runs in the explore role, which role-tools.ts already gates with --exclude-tools write,edit,multiedit (#238). Load-bearing rather than incidental: "Ask or Assume?" (69.4% on an underspecified SWE-bench variant) finds that an agent holding edit tools rationalises ambiguity away, because building is cheaper than asking.

Also fixes the #290 false positive I shipped yesterday

countEnumeratedFindings counts top-level - [ ], and in a /plan issue checkboxes are exclusive to ## Acceptance criteria. Measured on real bodies — #287→7, #288→6, #289→7, #366→6 — every count was test assertions, with zero of #287's five actual deliverables (**A.**–**E.**) seen, because bolded letters are invisible to it. A correctly-planned single-workstream issue therefore triggered a corrective re-dispatch essentially every time. Decomposition is now counted from the resolved deliverables[].

What this does NOT do

No change to /plan. Its job is to write the best spec it can; /work's job is to cope with whatever it is handed. Format markers stay a fast path, never a contract — and a reply with no ## Spec block falls through to the legacy verdict router rather than parking every cycle on drift.

Verification

Full §1 gate: build, tsc --noEmit, bun run check, 74/74 offline smoke tests. All touched files under the 500-line cap.

test-intent-resolution.ts — 30 assertions across the real range of spec quality: a well-formed spec; a one-line human bug report ("login is broken on mobile") that must park rather than invent deliverables; an issue contradicted by the code; already-implemented; too-large; a missing verdict; an unparseable verdict; and a reply with no ## Spec block at all.

The assertions that matter most are the negative ones — no deliverables are invented from a one-liner, and a proceed claimed alongside contradicting evidence is overridden.

One implementation note: sliceMarkdownSection matches exactly ## and cannot reach ### subsections. I added a local slicer rather than generalising the shared helper — parseWorkstreams depends on its current terminator behaviour, and widening a shared parser to serve one new caller is how subtle regressions get in.

Escape hatch PI_ENSEMBLE_INTENT=0. Docs: README knob, a troubleshooting.md section explaining each park reason, AGENTS.md §7.