fix(core): bound consumed job results

[jlongster]

Summary

Bound the process-global Job service's consumed terminal history to 25 entries, instead of keeping every completed shell/subagent job for the server's lifetime. Evict the whole record so its output, metadata, and completed Deferred snapshots can be collected.

This is a separate alternative to #46087, not stacked on it. The overlap was identified during implementation and the requester explicitly preferred a separate PR. This variant uses a smaller, count-only cache ordered by first consumption, removes acknowledged background records immediately, and keeps the patch limited to production code. It does not implement a UTF-8 output-byte budget.

Retention policy

• Keep running jobs and unconsumed terminal results: a job can finish before its caller gets from start() to block() / wait().
• A terminal foreground block(), explicit cancel(), or non-recoverable wait() makes a job eligible for the 25-entry history. Repeated reads do not refresh its position.
• Recoverable wait() alone does not consume the result, preserving the existing wait → background handoff, including immediately failed work.
• Pending notification IDs protect results from history eviction. completeBackground() removes the durable marker and the matching terminal in-memory record under the registry lock.
• Cleanup uses the existing generation token and notification ID, so an old observer/acknowledgment does not consume or remove a newer job with the same ID.
• Existing waiters retain their generation's Deferred/results. Future lookups of evicted history use the existing missing-result shapes.

No tool, public HTTP/protocol, database schema, configuration, or installed-binary changes. The separate Shell cache and transcript/output handling are unchanged.

Scope of the bound: this is a history-entry bound, not a hard cap on total Job or server memory. Running work, unconsumed handoffs, pending notifications, caller-held results, and the size of individual result/metadata objects are outside the 25-entry bound. A producer that never consumes its result still owns that handoff.

Evidence

An isolated installed opencode2 v0.0.0-dev-18911 shell-tool probe observed 32 completed Job records after warm-up, 132 after 100 further commands, and the same 132 after idle GC. All job scopes were closed; each record still referenced a fixture-tagged output node, while the separate Shell cache stayed at 25 entries. Session message history stayed empty.

• Snapshot counts and caveats
• Repeated 32-KiB shell-output workload

Those are pre-fix installed-server measurements, not a before/after memory claim for this patch. Bun's exported Map edges are incomplete, as documented in the report. Raw snapshots remain private.

Validation

• OpenCode Drive completed 40 concurrent mixed shell jobs, including normal output, nonzero exit, timeout, and 200 KB output cases; the Session continued after history eviction.
• The same Drive run completed a foreground subagent, a background shell notification, an interrupted shell, and a post-interruption shell recovery.
• Existing Job suite: 15 passed.
• Core bun typecheck, targeted Oxlint, both Effect AST rule sets, Prettier, and git diff --check passed.