feat: authorize coordination-layer dispatch in the scaffold

[radiusred-cody]

Closes #81.

Folds back finding 9 on #73: harness guardrails defer to AGENTS.md for sub-agent authorization, and the scaffold said nothing — so the M3 session correctly waited for the operator instead of dispatching the routed reviewer.

The scaffold's AGENTS.md now carries the conditional standing authorization (wording operator-confirmed 2026-08-25): if you are the operator's primary session — not dispatched as any role — dispatching a clean-context sub-agent as a routed App, via the identities doc's dispatch prompt, is authorized and expected. The same bullet carries the prohibition: a session dispatched as a role never dispatches another role, and never chooses or briefs its own judge. A platform-dispatched agent therefore reads a prohibition, not a licence — every Paperclip session fails the condition, preserving the implementer-never-dispatches invariant structurally.

identities.md states the tier gradient: solo is one session changing hats through the fixed dispatch template (the template is what keeps the hat-switch honest); a platform separates the hats into processes, watches the webhooks it registered at minting, and does all dispatching itself.

Scaffold test rides along per #46 (whitespace-normalized phrase assertions on the generated AGENTS.md).

Note for existing hubs: numberguess's AGENTS.md predates this and won't update itself — that propagation story is #85's virtual-_latest design.

[radiusred-testy]

Independent review (diff read before description; task #81 plan, milestone #67 M5-R1, and finding 9 on #73 checked; head extracted clean and go test ./... run green).

Verified:

• The scaffold bullet in agentsScaffold (internal/cli/init.go) is genuinely conditional: the authorization is scoped to "the operator's primary session — not dispatched as any specific role", and the same bullet carries the prohibition ("a session dispatched as a role never dispatches another role… never chooses or briefs its own judge"). A platform-dispatched agent reading this fails the condition and finds the prohibition in the same breath — the structure finding 9's remedy asked for, and the operator-confirmed conditional form (2026-08-25 note on #73).
• identities.md's new tier-gradient paragraph slots into the existing "Dispatching a role session" section coherently: it correctly points at "the fixed dispatch prompt below" (the numberguess-proven prompt is indeed below it), and the platform half — every session dispatched as a role, none meeting the condition — matches the plan's step 2 and the doc's existing webhook-dispatch text it follows.
• The scaffold test renders via scaffold() and asserts the load-bearing phrases whitespace-normalized, so reflowing the string won't break it. All three plan steps delivered, no undeclared deviations.
• Merge-simulated against the sibling PRs (#88, #89) touching identities.md: clean.

One consequence worth having on the record, and the PR body already carries it: existing hubs' committed AGENTS.md (this hub's included, and numberguess's) do not gain the bullet — propagation is explicitly deferred to #85. Until then, finding 9's guardrail stall can still recur on pre-existing scaffolds; that is a known, recorded boundary, not a defect of this change.