fix(agent): stop resume hydration from overfilling the context window

[tatoalo]

Problem

Resuming a long-lived cloud task can fail hard with Internal error: Prompt is too long, and once it does, every subsequent resume of that task fails the same way so the task is permanently stuck

Loosely:

1. Directory resume snapshots only cover the workspace, not ~/.claude, so the native session transcript is never on disk after a resume (warm: false). Every resume goes through hydrateSessionJsonl, which rebuilds a transcript from the task's ACP notification logs.
2. The logs endpoint concatenates the entire resume chain (oldest ancestor first), and in-session compactions are invisible in that log, so hydration resurrects history the session had already compacted away.
3. selectRecentTurns budgets 800k estimated tokens for 1M-context models at 4 chars/token. JSON-heavy tool payloads actually tokenize at ~2.5–3 chars/token, so the "800k" selection can be well over 1M real tokens. In the failing run, hydration kept 66 turns (~2.9M chars, estimated 732k tokens) that the API counted as 1,155,179 tokens against a 1,000,000 window. The 200k-window budget (150k estimated) has the same overflow, just less headroom to hide it.
4. The rebuilt transcript writes usage: 0 on every message, so the SDK's context accounting starts near zero: it can't auto-compact before the first API call (the run's compact_boundary reported preTokens: 16250 while getContextUsage() said 1.15M). The first call 400s, the SDK's recovery compaction can't help, and the turn fails as a non-recoverable agent_error.

Changes

• Estimate turn cost at 3 chars/token instead of 4, so JSON-heavy payloads are no longer undercounted
• Lower the hydration budgets to target roughly half the real window: 400k estimated for 1M-context models (worst case ~500k real), 80k for 200k-window models. This leaves room for system prompt, tools, skills, and the resumed run's own work, and un-bricks existing tasks (the next resume re-hydrates within budget)

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit ec17957.}

[github-actions]

Trivial constant-only fix with a clear rationale in comments; no logic, API, or data model changes.

[greptile-apps]

_{Reviews (1): Last reviewed commit: "chore(agent): state budget rationale ind..." | Re-trigger Greptile}