Post-v2.2: audit rad-con SESSION_RUNNER.md for content worth upstreaming #6

Context

During v2.2 Phase 1 field-testing, bin/sync --dry-run against rad-con surfaced that rad-con's local SESSION_RUNNER.md has diverged substantially from canonical starter-kit/SESSION_RUNNER.md:

This is real operator evolution — rad-con has ~90 sessions of accumulated discipline encoded into its runner.

Why this is a separate issue (not part of v2.2)

• Plan §9 (docs/planning/methodology-as-environment-plan.md) explicitly forbids content changes in v2.2. The scope was distribution (bin/sync, bin/status), not methodology content.
• Plan §2A + §5 Phase 8 exclude rad-con from the migration itself; this issue is about a different question: does canonical deserve an upgrade from what rad-con learned?

Upstream candidates (initial scan — not exhaustive)

Likely-universal improvements

• "Structural control, not a suggestion" phrasing for mandatory rules — with concrete session-collapse evidence (Sessions 62-63+ going from 9/10 → 1/10).
• Additional failure modes — verify whether canonical's 23 FMs already include rad-con's FM Proposal: render-dependency completeness discipline (silent font-fallback case) #12 (Workstream transfer amnesia), Phase 2.5 naming: consider Phase 2B for consistency with Phase 1B precedent #15 (Minimal handoff), CONTEXT_TEMPLATE.md references Claude-Code-specific 'auto-memory' — generalize for agent independence? #16 (False credit / fabrication), Cross-reference completeness — formalize the v2.6 self-review heuristic as a methodology Learning #19 (Plan-mode output is draft). If any are missing, they're good candidates.
• Grep-inventory Learning — "planning documents must search, not guess. Session 93 wrote a migration plan from architectural knowledge; Session 95 discovered 10+ missed files that 2 grep commands would have found."
• Plan-mode-as-draft Learning — "Plan-mode output is a draft, not a verified plan" (FM Cross-reference completeness — formalize the v2.6 self-review heuristic as a methodology Learning #19 / Session 98).
• Protocol-perishability Learning — "14 consecutive clean sessions collapsed to 1/10 within 12 hours of relaxed discipline" (Session 64).

rad-con-specific (do NOT upstream)

• Session-number references (B1 Phase 3: expand bin/sync to the full Option-B corpus (shared bin/_manifest.py) #34, Relicense the methodology under the MIT License #43, ..., feat(trim): SESSION_NOTES.md joins the ledger table; warn of an uncommitted front-matter edit (refs #93) #98) are rad-con history.
• Task mappings referencing docs/internal/PROFILE_DESIGN_PROMPT.md, docs/internal/AUDIT_PROMPT_TEMPLATE.md are rad-con's internal structure.
• Dated close-out entries embedded in the runner (e.g., "Mar 7, 2026 — Session 47: Phase 5").

Proposed workstream

1. Read rad-con's SESSION_RUNNER.md in full, side-by-side with canonical.
2. Classify each rad-con-only addition — universal / project-specific / already-canonical-in-spirit / noise.
3. Draft a content-change PR to canonical — structural-control phrasing, any missing failure modes, universal Learnings rows. Plan §9's "no content changes" constraint was scoped to v2.2; a post-v2.2 content release (v2.3?) is the right vehicle.
4. Keep rad-con's runner intact — this issue does NOT migrate rad-con. That's Phase 8 of the v2.2 plan, which remains "when the user says go."

Non-goals

• Migrating rad-con itself to ignored-mode (Phase 8 of v2.2 plan).
• Touching hamlib (also excluded per plan §9).
• Releasing this alongside v2.2 — ship v2.2 first, then consider a v2.3 content release.

References

• Plan: docs/planning/methodology-as-environment-plan.md
• rad-con runner: /Users/terrell/Documents/code/rad-con/SESSION_RUNNER.md
• Canonical: starter-kit/SESSION_RUNNER.md
• Field-test that surfaced this: commit a6bb8d6 (drift-safety patch)