refactor(core): advance sessions before running steps

[kitlangton]

Why

The Session runner distributes inbox scheduling across its drain loop, nested Step loop, and separate compaction/move probes. Its retry loop also carries input-promotion state even though promotion only happens before the first attempt.

What Changes

The drain now alternates between advancing the Session and executing a logical Step:

const next = yield* advanceToStep()
if (next._tag !== "Ready") return next
continuing = yield* runStep(next.context, step)

advanceToStep owns eligible control dispatch, instruction preparation, input promotion, and the first loaded context. runStep receives that context and a fixed logical step number; later attempts refresh context without delivering more input.

• Removes runSteps, eligible, the separate pending-compaction/move helpers, and the one-use Session lookup helper.
• Uses one control selection path instead of asking each control handler to inspect the inbox.
• Keeps queue policy in the runner rather than exposing it on Compaction.
• Removes 17 production lines and adds five characterization tests separately. No new scheduler service or framework.

Scheduling Contract

Situation	Preserved behavior
Initial instructions unavailable	Ordinary input remains pending; eligible controls can still run
Queued control at Location entry	May run before a carried model continuation
Follow-up steer cancelled during preparation	The Step stays steer-only; queued prompts and controls wait for proper dispatch
First Physical Attempt	Uses the context prepared by advanceToStep exactly once
Retry or overflow rebuild	Refreshes instructions/model/history without promoting pending input

Manual compaction delivery and start remain atomic. Its summary work remains interruptible and failure settlement stays protected. Movement returns to process-global execution to acquire the destination Location.

Scope

Core runner sequencing only. No queue/steer policy, public protocol, durable event, retry-budget, or manual-compaction interface change. Preserves the upstream background-recovery changes.

This is an alternative to #45335, not a stacked follow-up; the two approaches should not be combined. The independent interrupt/wait PR #45337 is unaffected.

Verification

# packages/core
bun run test test/session-runner.test.ts test/session-step.test.ts test/session-compaction.test.ts test/config/compaction.test.ts test/session-execution.test.ts test/session-run-coordinator.test.ts
bun run test test/session-runner.test.ts --test-name-pattern 'preflighting unavailable|Location entry before|refreshes preparation after|steer is cancelled during preparation|manual compaction|queued compaction|move|shares retry' --rerun-each 10
bun run test
bun typecheck
bun run build

# packages/sdk (isolated configuration)
bun run ../core/script/test.ts

• Full Core suite: 2,341 passed, 16 skipped, no failures.
• Focused runner/attempt/compaction/execution/coordinator coverage: 239 passed.
• Repeated scheduling and recovery cases: 220 passed across ten repetitions.
• Isolated SDK suite: 24 passed.
• Core typecheck and build passed; workspace pre-push typecheck passed all 32 tasks.
• Formatting, Effect simplification checks, and diff whitespace checks passed. Oxlint reported no errors and three pre-existing warnings in the test file.
• The cancelled-steer regression test failed on the first refactor cut and passed after preserving the follow-up delivery scope. The move variant also verifies transport closure and durable movement, not merely inbox consumption.