Avoid empty managed Runtime scan intervals

[SaladDay]

Managed Runtime scans currently spend an entire five-second ticker interval at allocation EOF without observing any allocation. A small deployment therefore observes its instances only every other tick, adding unnecessary wait to wake requests and cleanup.

On EOF after a nonempty cursor, advance the existing bounded initialization operation and read the first page in the same reconciliation call. Preserve the five-second ticker, at most 32 allocation observations per call, one initialization operation per full scan, and a single refill so an empty store cannot spin. Provider APIs, harness behavior and suspension eligibility are unchanged.

Validation:

• Full make check passed on Linux with real PostgreSQL; GitHub check passed.
• New 0/1/31/32/33/65-allocation, EOF cleanup, cancellation and empty-store regressions pass; the nonempty scan and cleanup regressions fail against the main baseline.
• Initialization/restart/uncertain-outcome tests now cover 33 allocations and require all of them to be serviced between initialization operations.
• A fresh independent full-diff review found no substantiated in-scope issues. The reviewer independently passed managed/runtime store regressions, execution tests and both provider packages.
• Four real Core/Codex/KVM runs using Kimi K3 passed: two baseline and two candidate samples, each with two Turns, idle source-VM release, exact snapshot restore, retained history/files/configuration, one-time initialization, equal idempotent retries and owned cleanup.

Measured request submission to Core Turn.started, using one server clock and PostgreSQL timestamps:

Path	Baseline samples	Candidate samples
New Session	25.146 s, 26.409 s	25.132 s, 26.376 s
Next Turn after suspension	17.415 s, 17.858 s	12.465 s, 12.937 s

Resume improved by 4.951 s and 4.921 s in these pairs. New-Session startup showed no meaningful improvement. The runs reused the same cached immutable image, helper and native runtime; image download, native first-tool timing and model response completion are outside this measurement. Two samples per version do not establish a percentile or SLA.

The sessions compatibility job passed persistence tests, then reproduced existing CI-001 at official_agent_references.py:73 (Expected BadRequestError). Its later container stage did not run. The same stale assertion was recorded before this change; no assertion or gate was bypassed.

Plan and evidence: Document A, PERF-001.