Implement design_docs/006-the-coders-front-door.md nf-6lm ← Beads

Twinned from https://github.com/cfsh/ninefold/issues/768 by tools/beads/import_github.py's reconcile pass.

Implements [design_docs/006-the-coders-front-door.md](https://github.com/cfsh/ninefold/blob/main/design_docs/006-the-coders-front-door.md) — the design round for #733. Doc PR: #767.

The recommendation, in two lines. The workflow is the output of a generating process: an incident produces a paragraph in an always-loaded file instead of a refusal in a tool, and CLAUDE.md regrew 67% in the three days after the last de-bulking. Split the coder's manual from the operator's, give the coder one front door (./dev.py) that owns every session-specific value, delete the dev's poll looping outright — @aedanpope, 2026-08-09: *"lets delete all the poll looping first, and only add it back when we need it. I can rely on CTO's own loop in the interim"* — with the box-level backstop a contingency, not a prerequisite, and budget the always-loaded file so the next incident has to buy its way in.

⚠️ Read design_docs/006 §8.1 first — the execution order. It is argued from docs/DEV_PRINCIPLES.md and runs deletions → cheap fixes that stop an active loss → class-killers → the front door → contingent. §8.2's table is the dependency truth and wins where the two conflict. The largest wins are Wave 0 and are pure deletion with no prerequisites, so a session that only ever does Wave 0 has still done the most valuable part.

Shape of the work. 22 pieces in the doc's §8, roughly 10–14 PRs. They are not all one stack:

• Land first, independent, no dependencies — S1a (the operator slice of the manual split: board, CLAIM/RELEASE, reaping, dispatcher and f5 health) and the small principles-lens fixes. ⚠️ S1 is three pieces, not one, and only S1a is dependency-free. S1a′ (the poll loop's own prose) waits on S5 — 189da67 is @aedanpope reversing exactly that move 46 minutes after the 2026-08-05 split, because while a coder runs the loop it is the loop, not reference. S1b (coder-facing prose in notes/) waits on the refusal that replaces it (#765).
• A single stack over dev.py — --json on the consumed plumbing → skeleton → pr (+ 19a, opening the 2nd/3rd/4th stacked PR — this is E4's mandated split made cheap, and it needs nothing else) → status (queue-shaped: every PR, ready-set first) → transactional stack merge for any member, since Tier-1 means merge order is whatever goes green first. ⚠️ Retiring the per-session cron is NOT in this stack — it is piece 11, land-first alongside 1a with no prerequisites, per the 2026-08-09 decision. dev.py wait and the box watchdog are contingent and may never be built. ⚠️ The cron retirement is the only irreversible piece and must migrate per session, opt-in, never as a flag day — several sessions are live on this path at any time. It lands early, but it still lands per session.
• ⚠️ dev.py feedback (S16, piece 22) — the channel that stops the tracker being the only place a dev can put a symptom. Standing review issue #771 is its guard; design 006 §5.10 says that if #771 goes stale, close S16 rather than leave a channel absorbing reports. S16 also proposes D4 for docs/DEV_PRINCIPLES.md, offered for @aedanpope rather than applied — do not land it without him.
• Independent of the front door entirely — pieces 14–18, which came out of the session testimony rather than the cognitive-load argument. Several are one-liners against live bugs, and #728 is a rule in CLAUDE.md that is provably impossible to comply with (5 sessions hit it). If this doc is only half-implemented, these are the cheapest harm removed per unit of work.

Nothing needs @aedanpope. ✅ All three questions are resolved: §7.6 by measurement — ruleset 20540766 is enforcement: active as of 2026-08-09, so there is no decision, and the live consequence is #763 (piece 8) rather than a governance call; §7.3 is answered (delete first, rebuild on demonstrated need) and §7.2 is resolved by default — his S7a point (the existing 5-line refusal should teach the trade) dissolves the number question, and S7b is deferred: it is a second gate for an anticipated need, so B4 says ship S7a, measure, and build S7b only if CLAUDE.md grows anyway. ⚠️ An implementing session should not set a budget number on its own authority — §5.4 objection 3 is that the budget mostly binds @aedanpope.

⚠️ The work breakdown is deliberately in the doc rather than filed as separate issues (.claude/commands/design.md §6; the precedent design 005 set with #686). Whoever takes this chunk may split it then, holding the whole thing and in the right order.

Unclaimed and unlabelled — backlog, for a triage round.