refactor(prose): skill-handoff convergence — one grammar, bridge plan integrity

[leeovery]

Standalone PR off main (independent of the batching stack). Fixes the Portal stall (conclude-discovery printing its bridge payload and stopping) as one instance of a family, per the plan agreed with Lee.

The mechanisms, made explicit (CONVENTIONS: "Loading, Invoking, and the Bridge")

• Loading a reference reads into context; variables resolve because prose stated them first.
• Invoking a skill is a Skill tool call at a boundary, enriching the running context — nothing is cleared. Canonical form: imperative before the payload fence ("Invoke the workflow-x skill (Skill tool) with the next fenced block as its arguments…"), fence as pure payload, terminal after — no STOP, no after-fence instruction (too late), no in-fence instruction (gets printed, not executed).
• The bridge owns the only context-clear, phase→phase via plan mode.

The sweep (34 sites, 24 files)

• In-fence Invoke the workflow-X skill. payload tails removed everywhere (conclude-* ×9 incl. review-actions ×5, entry invoke-skill ×8 skills, spec-entry handoffs ×6).
• conclude-discovery gains the missing imperative and loses the terminal STOP that licensed the stall.
• Entry files' imperatives move from after the fence to before it; payload-construction notes move above too (command-prelude).

Bridge plan integrity (the ~10% contamination)

All 8 plan-template sites: "verbatim — this is the complete plan", plan mode's investigate/verify/design job explicitly disclaimed, nothing learned this session added (the next context is designed to start empty), ## User instructions as the sole sanctioned addition — and ## How to proceed now addresses the human plainly: approve with "Clear context and continue" (the option migration 034 keeps enabled).

Test plan

• Lint check 14 (buried invocation imperative) + fixtures; corpus clean.
• npm test 1560/1560 · conventions lint 29/29 · typecheck clean.

🤖 Generated with Claude Code