fix: persist modelContextLimit, skip min-threshold when unknown, percentage defaults, keep blocks active across restarts

[Neko-Yukari]

Summary

Four related fixes for premature compression reminders on modern large-context models (256K-1M tokens), observed as "restart then immediately asked to compress" on a ~300K / 1M context.

1. Persist modelContextLimit (state/persistence.ts, state/state.ts)

opencode runs the chat.message transform before the system.prompt hook, so on the first message after a restart state.modelContextLimit is not yet cached and is not persisted. resolveContextTokenLimit then fails for percentage limits. Persisting the value restores it before the first transform.

2. Skip min-threshold check when limit is unknown (messages/inject/utils.ts)

Previously overMinLimit = minContextLimit === undefined ? true — with the limit unresolvable, a normal ~300K context on a 1M model triggered compression nudges on every turn. Now it skips the check (false); the real model limit is still enforced by the API layer. (max side already behaved this way.)

3. Default limits as percentages (config.ts)

Absolute 100000/50000 defaults fire far too early on 256K-1M models. Defaults are now 85%/60% of the model context, falling back gracefully when the limit is unknown.

4. Keep compress blocks active across restarts (messages/sync.ts)

compressMessageId (the assistant message executing compress) is marked ignored/synthetic and never persisted, so after a restart syncCompressionBlocks deactivated the block and the full original context was re-injected. When the origin message is missing but the anchor message still exists, the block is kept active so the summary keeps being injected.

Tests

109/109 pass. Added:

• min-threshold skip when modelContextLimit unknown
• 300K-context / 1M-model no-false-alarm regression
• modelContextLimit persistence round-trip
• sync-block keep-active + summary-injection regressions