…w goals (NousResearch#34196, NousResearch#34197)

Two related /goal bugs:

(review/reflect/suggest/analyze) unless the assistant uses a magic
phrase like 'goal complete'. The synthetic continuation loop escalates
reflection into producing concrete artifacts that the goal only listed
as *examples* of possible help.

untracked` even when the user did not ask for staging/commit/push.
This races with preflight compression and survives session split,
turning a scoped 'done' answer into out-of-scope artifact production.

Both bugs converge on the goal-judge prompt machinery in
`hermes_cli/goals.py`. The fix is layered, minimal, and reviewable:

1. Tighten JUDGE_SYSTEM_PROMPT with three new explicit guardrails:
   - EXPLORATORY goals (review/reflect/suggest/analyze) are completable
     by a substantive synthesis — do NOT require additional artifacts.
   - Do NOT infer incompletion from untracked / unstaged / uncommitted
     files unless the goal explicitly required staging/commit/push.
   - Do NOT require a magic phrase like 'goal complete'.
   - Treat 'for example' / 'maybe' / 'you could' items as illustrative,
     NOT as required deliverables.
   - Scope-narrow goals (one file, one section, one specific change)
     are DONE when that exact scope is confirmed done — do not expand.

2. Add a transparent keyword classifier `_classify_goal_shape(goal)`
   that returns 'exploratory', 'illustrative', or 'concrete'. Cheap,
   reviewer-friendly substring detection — the LLM judge still makes
   the final DONE/CONTINUE call, but it now sees what kind of goal it
   is. Kept intentionally simple so behaviour is easy to audit and
   tune from issue feedback.

3. Append a corresponding goal-shape hint to the user-prompt template
   when the goal is exploratory or illustrative. The hint reminds the
   judge that for those shapes, a high-quality synthesis IS the
   deliverable. Concrete goals get the original strict template
   unchanged.

4. The with-subgoals template (already enforces strict per-criterion
   evidence) deliberately does NOT receive the shape hint — the
   user's explicit /subgoal criteria take precedence over goal-shape
   heuristics.

Why prompt-level vs adding new state:

This intentionally avoids adding new GoalState fields, new gateway
plumbing, or new compression-lifecycle coupling. The goal judge is the
single point where 'should we continue?' is decided; teaching it to
read goal shape correctly is the smallest change that addresses both
issues' root cause without touching the compression race window
described in NousResearch#34197 #2-#5. Those lifecycle concerns are real and
documented in the issue's 'Proposed fixes #3-#6' — they belong in a
separate gateway-side change. This PR fixes the judge's bad
'continue' verdict that triggers the lifecycle problem in the first
place. Without the bad verdict, the race window in NousResearch#34197 has nothing
to race over.

Tests (17 new, all 67 in test_goals.py pass):
- TestClassifyGoalShape (9 tests): exploratory/illustrative/concrete
  classification including the sanitized NousResearch#34196 repro goal text.
- TestJudgeSystemPromptGuardrails (4 tests): system prompt mentions
  exploratory goals, warns about untracked files, warns about
  requiring magic phrases, warns about illustrative examples.
- TestJudgePromptIncludesGoalShapeHint (4 tests): the user prompt
  receives the shape hint for exploratory/illustrative goals, does
  NOT for concrete goals, and the with-subgoals template skips the
  hint to preserve its strict per-criterion evidence rule.

Refs: NousResearch#34196 NousResearch#34197
Closes: NousResearch#34196 NousResearch#34197

Co-authored-by: Cursor <cursoragent@cursor.com>