perf(platform): let staged skills survive a sandbox restart unsent

[yannickmonney]

Every turn after a session's idle stop re-sent every equipped skill bundle inline (up to 32 MiB each), because the reuse probe identified files by digest but did not send the digest, and the runtime's record of what it staged dies with the container.

Changes

perf(platform): let staged skills survive a sandbox restart unsent

Skill bundles stage as inline files with a reuse probe that named each
file by the digest of its bytes but did not send the digest. The runtime
could confirm a hit only from its in-memory record of what it had staged,
which every restart of a stopped session (5 minutes after a turn) loses,
so each later turn sent every bundle again inline: up to 32 MiB per skill,
read, base64-encoded and carried platform -> spawner -> runnerd at 1.33x.

The probe now carries each file's sha256, which the runtime already checks
against the file on disk when an item names one, so an unchanged bundle is
confirmed without its bytes after a restart too.

Verification

• vitest: session_client.staging.test.ts + session_client.test.ts 101 pass (the probe test now asserts each file's digest).
• @tale/platform typecheck; type-aware oxlint clean on the touched files.
• The runtime side is unchanged: file-ops.ts already verifies sha256 against the file on disk (held by file-ops.test.ts).

Part of the sandbox footprint and robustness pass (#4533, #4538 and siblings).