orchestration: split the Orchestrator — plan → render → deliver, a break becomes an inspectable value

[genwave-radio]

🎛️ Finding (August 2026 architecture analysis, §5)

The Orchestrator (~1,570 lines) trends toward god-object territory. The pick pipeline — request rung → persona rung → envelope query → trust-but-verify → degradation ladder — is a coherent standalone policy; extract it (working name MusicSelectionPolicy), leaving the Orchestrator to own cadence and unit assembly.

⏰ Timing — sharpened from "judgment call" to a scheduling rule

The analysis says extract when a second consumer appears. Project #3's cycle IS that moment: #300 (boundary-fit bug), #320 (straddle handoff), #383 (dayparting) all churn exactly this pipeline. This extraction is the opening task of project #3's first /plan — before the churn, not mid-epic (the wardrobe epic's "data before law" rung discipline).

📎 Scope note

Pure Orchestration-internal restructure — no seam changes, no Host contact; the existing Orchestration.Tests suite is the harness.

[genwave-radio]

Repository review 2026-09-10 (R07), re-checked on v5.8.2. Current numbers: 2,687 lines, ~290 of them header comment before the declaration at line 290; 27 constructor parameters, 15 optional. Growth: ~1,570 when this issue was filed → 2,536 at v5.4.0 → 2,688 at v5.6.0 → flat since (the last three epics lived in Ads and Host). The 'opening task of project #3's first /plan' trigger has not fired. The review asks for more than the music-pick extraction: a coherent planning snapshot and an inspectable break plan (ordering, speaker context, deadlines, delivery identity), then rendering and delivery extracted incrementally, carrying gh-#772's speaker snapshot through. Sequence: characterise the audible ordering contract first, then extract; never as one architecture-cleanup PR. Second and third candidates, separately justified: MediaRepository (2,419 lines), LlmCopyWriter (1,553).

[genwave-radio]

✅ Shipped — v5.10.0 and v5.11.0

Closing this as the Orchestrator split, which is not quite what the title said when it was
filed. Worth recording the difference, because the August ask and the thing that closed it are two
different pieces of work.

📜 What happened to the original scope

This issue was filed in August asking for one extraction: MusicSelectionPolicy, the pick pipeline.
That shipped on 2026-08-31 — commit 5313232 (T218 / STORY-295), 928 lines out of the
Orchestrator, three weeks before this epic began.

What kept the issue open was the 2026-09-15 repo-review comment (R07), which
re-scoped it: "the review asks for more than the music-pick extraction: a coherent planning
snapshot and an inspectable break plan (ordering, speaker context, deadlines, delivery identity),
then rendering and delivery." That re-scope became SPEC F184–F194, STORY-451…461, and is what
this closes on. The title was updated to match before closing.

🧩 The split

Orchestrator.EnqueuePatterAsync decided, claimed, rendered and enqueued in one interleaved pass,
so the only way to see what a break contained was to listen to it. It is now four collaborators and
a value:

Class	Lines	Owns
BreakPlanner (+3 partials)	726	Decides the break — slots, ordering, reservations, speaker snapshots. Renders nothing.
BreakRenderer	135	Plan in, one SlotOutcome per slot out. No logging, no events, no buffer.
BreakDelivery	251	Buffer order, drop logs/events, reservation outcomes.
HandoffCeremonyProducer	549	Arms, holds and enriches the two-piece SignOff/SignOn ceremony.

BreakPlan is an immutable record — ordered slots, one speaker snapshot per spoken slot, one named
reservation per claimed thing — with ToTrace() for diagnostics. A test can now inspect a break
with no audio and no engine.

Orchestrator.cs: 2,687 → 1,098 lines (1,055 + a 43-line Orchestrator.Render.cs partial), a 59%
cut against the 2026-09-17 baseline.

🔊 The one audible change

gh-#772 (shipped in v5.10.0): every spoken slot carries the speaker it was planned for, not
whoever happens to be on-air when the render finally runs. SpeakerSnapshot is new on
GenWave.Abstractions; the TTS cache keys on the snapshot's own content hash, so a card edit after
planning cannot be served stale.

🚢 The PRs

PR	What
#805	OrchestratorBuilder front door, root NoOp resolution, the frozen characterisation replay
#808	The break contract pinned cell by cell
#818	The break plan is a value; BreakPlanner decides the break
#820	The speaker travels with the plan (gh-#772)
#821	v5.10.0 pins; Abstractions surface baseline → 5.10.0
#824	Extract HandoffCeremonyProducer
#825	Extract BreakRenderer
#826	BreakDelivery + the Orchestrator's fitness pins
#829	v5.11.0 pins + the three F193 rulings

Every PR ran the full solution green with an adversarial review gate per task. The characterisation
replay (Story452_BreakCharacterisationReplay.cs, Story452_PinnedTables.cs) was frozen at the
start and stayed byte-identical throughout — exactly one authorised touch, T522, which added a
trace fact and deleted no assertion.

🔍 Two things stated honestly rather than quietly

F192.5 is a named deviation, half-wired. A ct that fires before delivery now means no item of
the unit reaches the buffer at all, closing a partial-enqueue gap observable only at shutdown. The
buffer-side guarantee is live. BreakDelivery.Abandon — the reservation-side half — is pinned by
its own fact but has no production caller: RenderPlanAsync returns before delivery is reached.
Wiring reservation settlement end-to-end is a later story's job, and the changelog and SPEC both say
so rather than claiming the whole guarantee.

One fitness pin ships baselined, not clean. SPEC F193.2 claimed no class in
GenWave.Orchestration declares an optional constructor seam. That is false for 10 classes / 25
parameter pairs, so the pin ships as a named, dated baseline asserted by set equality — a new
violation fails, and so does a fixed violation whose entry is not struck in the same change, so it
cannot fail-open while it burns down.

🔭 Spun off

• Burn the optional-seam baseline in GenWave.Orchestration down to zero (SPEC F193.2) #827 — burn the optional-seam baseline to zero: 25 pairs across 229 call sites, four PRs.
• Make the one-construction-path pin real: migrate 7 spec helpers, then scan IL (SPEC F193.4) #828 — make the one-construction-path pin real: migrate the 7 spec helpers that build an
  Orchestrator by target-typed new(…) (invisible to a text scan — 9 real sites against the 2 the
  scan sees), then replace the scan with an IL call-site scan.

#828's phase 1 is #827's prerequisite; doing it once buys both.

📦 Released

v5.11.0 — https://github.com/GenWave-Org/genwave/releases/tag/v5.11.0 — stack gate green, demo
box pulled and on the tag.