stella-serve: TurnTally cannot distinguish a parked turn from a wedged one — the exact question it exists to answer

[macanderson]

Problem

TallyFold::observe (crates/stella-serve/src/observe/tally.rs, the match event around line 51) drops AgentEvent::TurnParked and AgentEvent::TurnWoken into its inert _ => {} arm.

That is precisely the distinction the tally exists to make. Its own field doc for TurnTally::stages (crates/stella-serve/src/observe/event.rs, around line 535) says:

A turn whose stages stopped advancing while a reverse request's wait climbs is wedged; one that keeps advancing is merely slow.

A parked turn (#1471) stops advancing stages on purpose — it is probing external state on the engine's own clock, with zero model calls, for up to its deadline. With park/wake dropped, a host reading the tally sees a stages-stall it cannot tell apart from a hang, and the one signal that would explain it was on the stream and discarded.

This is not a regression from #1857 — before the typed events the park arrived as AgentEvent::Text, which this fold also ignored. It is a gap that only became fixable once the events existed.

What done looks like

• TurnTally gains park accounting — at minimum a parked_spans count; polls_used and/or accumulated parked seconds are worth considering, since "wedged vs deliberately waiting for 20 minutes" is a duration question.
• TallyFold::observe gains explicit TurnParked/TurnWoken arms. Follow the existing saturating_add idiom used by every other counter in that match.
• A host reading the tally can answer "is this turn wedged?" correctly for a parked turn. Assert it: a fold over [Stage, TurnParked, TurnWoken, Stage] must be distinguishable from a fold over [Stage] alone.
• Check whether TurnTally is a wire type reaching docs/wire/serveframe.schema.json. If so, regenerate via scripts/export-agentevent-schema.sh in the same PR (the wire-schema gate) and keep the addition serde-additive so a host recorded before this parses.

Constraints

• Content-free. The tally is observability. TurnParked.description is tool-authored free text and must NOT be counted into or copied onto the tally — counts and the closed WakeReason token only. Same rule the bus payloads follow in crates/stella-core/src/driver/lifecycle.rs::turn_parked_payload, and the diag bridge in crates/stella-cli/src/diag_bridge.rs.
• The existing test crates/stella-serve/src/observe/tally.rs (fold.finish() == TurnTally::default() for content events) pins that text/reasoning/steer contribute nothing — do not weaken it.
• Park semantics must not change; this is an observability gap only.

Context

• The events landed in feat(stella-protocol): typed TurnParked/TurnWoken wire events and a deck-rendered park state (#1857) #1994, closing Parked waits: dedicated wire events (TurnParked/TurnWoken) and a TUI heartbeat instead of synthetic Text #1857.
• Consumer sweep in fix: unbreak main's fmt gate, and stop the parked span vanishing from the observatory journal and fleet row (#1857 follow-up) #2005 fixed the two regressions this same wildcard problem caused (the observatory journal allowlist and the fleet dashboard row) and deliberately left this one, because it needs a TurnTally field rather than a match arm.
• Related: stella-serve: RemoteToolExecutor cannot forward parked-wait requests across the wire #1858 (RemoteToolExecutor cannot forward parked-wait requests across the wire) — a served turn cannot park at all yet, so this is the observability half of the same story and the two are worth sequencing together.

Refs #1857, #1471