docs(plan-template): track PLAN_TEMPLATE.md + adopt falsifiable-verify & blast-radius hardening

[lantisprime]

Summary

Tracks docs/PLAN_TEMPLATE.md (it was untracked) and adopts the evolved upstream pi-extensions Appendix B so a lower-capability LLM can execute a plan without silent drift. Net: a weak executor now gets one editable file per step, a labeled action kind, a verbatim anchor, and a verify that is required to fail on a stub — closing the three places weak models drift (multi-file edits, vague anchors, hollow greens).

What was adopted (user-approved A–F)

Item	What	Where
A	Falsifiable Verify — parent rule + deny-list (tolerant test $?, self-fulfilling grep, happy-path-only) + positive obligation (name observed + expected value) + compound-bash caveat	new §A.6b
B	UNGUARDED-IN-CI tag — a MUST/Safety row covered only by a skippable smoke is tagged + names the manual step (no hollow green)	§4, §7
C	No aspirational output — a printed "checking X" must assert X	§A.2 item 9, §14
D	Test-body rigor = code rigor — CREATE tests with full verbatim assertions over real captured output	§A.7
E	One-file-per-step + CREATE/EDIT/APPEND + executor-ready gate	§A.2, §A.6, §A.7
F	Blast-radius patterns — red-then-green, pure-extraction-first, thin-wrapper, fixture-change ledger, discriminating sentinel, flag-high-blast-radius, mock-project E2E, non-deterministic-guard	new §A.9

Plus 6 rows in §20 mapping each new rule to its enforcing section (our sync convention).

Adaptation notes

• Re-grounded examples in .mjs/bash and real slices (P3a pure-extraction; the ESC cross-session unmask as the fixture-change-ledger cautionary tale).
• Compound-bash caveat makes the negative-control pattern legal under this repo's compound-bash-gate.
• Dropped pi's LLM-app specifics (--append-system-prompt, --mode json); kept only the transferable non-deterministic-guard logic.
• Preserves all existing governance (RFC R-grounding, Rule 12 token budget, three-layer review, lessons table).

Scope

Docs-only. No hooks/scripts/patterns touched. 524 → 655 lines (one new tracked file).

🤖 Generated with Claude Code

[lantiscooperdev]

Bot review — pre-merge coherence pass (not an approval)

Docs-only change; reviewed for internal consistency and non-regression of existing governance.

Verified

• Section ordering intact: §A.6 → §A.6b → §A.7 → §A.8 → §A.9; the 20 numbered sections + Appendix A are all present and ordered (grep over headers).
• Cross-references resolve: §A.6b ↔ §4/§7/§A.9, §A.7 ↔ §A.6/§A.6b/§A.5, and the 6 new §20 rows ↔ their enforcing sections all point at existing headers.
• No governance regression: RFC R-grounding (§4), Rule 12 token budget (§6), symlink matrix (§7), three-layer review (§19), and the lessons table (§20) are untouched — this is additive hardening of Appendix A only.
• Repo-fit: the compound-bash caveat in §A.6b keeps the red-then-green negative-control pattern legal under compound-bash-gate; examples are re-grounded in .mjs/bash (P3a, ESC) rather than pi's LLM-app specifics.

Note (non-blocking)

• §A.6b's "greppable Verify-cell lint" is described but not yet a CI check. If we want the falsifiable-verify rule enforced rather than documented (Rule 13), a follow-up could add a lint over plan step-tables. Out of scope for a template change.

Verdict

ACCEPT (docs). No blockers. Recommend merge after human approval.