…plate

Address PR review feedback (#822)

The inlined solo template dropped the "(You can also redirect to
/ce-brainstorm if this is bigger than you initially thought...)"
parenthetical that the reference template carries. Inlining was
supposed to make the gate fire reliably without the reference loading,
so dropping the escape-hatch line from the inline copy weakened the
guardrail it was meant to preserve. Sync the inline template back to
the reference's wording.

Apply the shape and discipline changes from ce-brainstorm's
scoping-synthesis fix (#829) to ce-plan's Phase 0.7 / 5.1.5:

- Tier guard on auto-proceed: Lightweight + zero call-outs is the only
  path that skips the confirmation gate. Standard and Deep plans always
  fire the confirmation gate even with zero call-outs, because substance
  earns the checkpoint. A 1-3 line summary on a Deep plan is exactly
  the rubber-stamping case the gate is supposed to prevent.
- Confirmation phrasing names what happens on confirm ("Confirm and
  I'll proceed to research, drawing on this scope" / "Confirm and
  I'll write the plan next..."), replacing the ambiguous "Confirm
  to proceed."
- Detail test for each surviving call-out and summary bullet: 1-2
  lines max, conversational not documentary. The count cap was
  gameable without it -- three call-outs could each be a 6-line
  paragraph and still "fit."
- Re-cut rule extended to fire on detail overflow, not just count
  overflow.
- Summary form is flexible: prose, bullets, or mix, whichever
  communicates best. Tier-aware budgets (Lightweight 1-3 lines;
  Standard 3-5 lines or 2-4 bullets; Deep 4-6 lines or 3-6 bullets).
- Rename "Scope Summary" / "Synthesis Summary" to "Scoping Synthesis"
  for parity with ce-brainstorm's terminology.
- Soft-cut option wording updated per the parity note in #819 (the
  "redirect" verb collided with the unrelated self-redirect mechanism).

Skill doc updated -- the Quick Example referenced "short prose
summary" and "the gate skips when there are no forks worth flagging,"
both of which would mislead a reader under the new behavior.

…scipline

A test run of the new two-paragraph synthesis still surfaced
plan-pitch leakage in three patterns the rules didn't yet block:

1. The agent claimed "one PR" — a sequencing decision plan-write
   produces, not something knowable at synthesis time.
2. "Plan-specific scoping" was enumerating where the implementation
   reaches into the codebase (file paths, Implementation Unit
   inventory) instead of stating scope-claim decisions.
3. Call-outs kept the 3-5 line "name fork, explain A, explain B,
   my default is X" rationale-dump shape, which is exactly what
   belongs in Key Technical Decisions in the plan body.

Encode the underlying rule explicitly: the synthesis is composed
before plan-write, so it can only surface what the agent knows from
the brainstorm + research + posture commitments. Implementation Unit
boundaries, PR count, commit/branch sequencing, effort estimates,
and exact file paths are all plan-write outputs the synthesis cannot
honestly claim. Even when the agent has formed plan-write opinions
earlier in the session, those stay internal until plan-write.

Other refinements:

- Reword plan-specific scoping from "what this plan covers vs defers
  vs expands" to "scope-level decisions" — the "covers" framing was
  pulling agents toward inventory.
- Make plan-specific scoping items pass the same affirmability test
  as call-outs: the user can affirm or redirect without reading code.
- Strengthen the call-out template placeholder to forbid
  multi-sentence rationale and "my default is X" pitches.
- Generalize the bare-ID anti-pattern in source-vocabulary discipline
  (AE4, R6, F3 all flip the user back to the brainstorm).

A test run showed two wording patterns the prior rules didn't block:

1. Bare ID references resurfaced ("AE1-AE3", "AE4", "AE5", "R6")
   even when the cases were already named in plain terms in the same
   sentence. Strengthen the source-vocabulary rule into a mechanical
   pre-emit scan: before emitting, look for `AE\d+`, `R\d+`, `F\d+`,
   `A\d+`, `U\d+` patterns and replace with plain names.
2. Numerical attestation ("all nine requirements, all three flows,
   all five acceptance examples") read as the agent showing its
   work — "covers the full brainstorm scope" already conveys the
   claim and the count adds nothing the user can affirm. Add as a
   named anti-pattern alongside synthesis-as-plan-pitch.

Both are wording-polish refinements on top of structural rules that
are now landing. Reference-only changes; no SKILL.md inline updates
needed since these refine quality, not gate firing.

Fresh-session test still produced "The touch surface is X (subpaths), Y
(subpaths), Z..." enumeration in paragraph 2 plus "all 9 requirements"
numerical attestation. The rules forbidding both were in the reference
but the touch-surface prohibition was buried in a comma-separated list
of NOTs and the file-path scan didn't exist yet.

Promote both to load-reliable inline placement in SKILL.md and add the
file-path scan as a pre-emit mechanical check:

- "Do NOT enumerate the touch surface" gets its own bold inline
  paragraph in both Phase 0.7 and Phase 5.1.5. Names trigger phrases
  ("The touch surface is...", "This plan touches...", "The
  implementation reaches into...", "Files modified include...") so the
  agent recognizes the pattern even when the buried rule misses.
- Pre-emit scan rule expanded from bare-IDs-only to bare-IDs + file
  paths. Same mechanical shape: before emitting, scan for `path/like.md`
  / `path/like.py` patterns and cut unless the path IS the topic of an
  explicit fork in the call-outs.
- Reference section reorganized: source-vocabulary covers vocab choice;
  a separate "Pre-emit mechanical checks" bullet groups both scans
  with examples of allowed vs forbidden path usage.