feat(compaction): adaptive failure backoff and actionable context budget errors

[ChrAlpha]

What this is

Two halves of the same failure experience, landed as two commits: when compaction fails under pressure the retry cadence now adapts (commit 1), and when a turn is ultimately rejected for context budget the error tells the user what to do about it (commit 2). The second half was originally #1123, folded in here since both exist to make the context-budget failure path recoverable and legible.

How it works

Cooldown backoff (feat(compaction): retry on exponential backoff under hard context pressure)

The compaction failure cooldown was a flat five minutes regardless of pressure, so one transient summarizer failure at the blocking threshold left every turn degraded or failing for the full window.

• Failures now track consecutive attempts per session. Automatic triggers fired at or above the hard threshold (75% of the context window) pass HardPressure — both the async trigger and the sync backstop compute it from the same pressure/budget pair — and re-attempt on an exponential schedule: 30s, 1m, 2m, 4m, capped at 4m30s.
• The cap sits strictly below the 5m entry-expiry window, so a capped retry still finds the failure entry alive and attempts keep accumulating instead of resetting through expiry. Steady state under a permanently failing summarizer with sustained hard pressure is one attempt per 4m30s — effectively the old flat cadence, but a transient failure recovers after 30s instead of five minutes.
• Soft-pressure triggers keep the flat cooldown; success still clears the state; manual runs still bypass; the in-flight-owner attach still precedes the cooldown check.

Actionable budget errors (feat(apperror): make context budget errors actionable)

A turn rejected for context budget surfaced only "Required context exceeds the model context budget." — accurate, but with nothing the user can do about it, in English, on every surface.

• The Detail strings for context.protected_overflow and context.budget_unsatisfied now name the recovery paths: run /compact to summarize older history, shorten the request (window-too-small only), or switch to a model with a larger context window. Channel replies use the Detail directly; the web UI localizes by code.
• All three web locales updated per the codesync(error-catalog) contract; pinned expectations updated in Go tests and the web vitest suite.
• Reviewed: a repo-wide sweep found no other consumer of the old strings; /compact is registered on the shared channel command group and shipped as a web chat quick action; channel error rendering applies secret redaction only, no length caps. For a reserve-driven budget_unsatisfied (window smaller than the output reserve) only the model switch helps — the wording hedges all three remedies.

Verification

• Unit tests pin the exact backoff timeline: early hard-pressure retry after the first step, doubling on repeat failure, cap-below-expiry with attempts persisting across a capped retry, reset after success, and a full RunCompactionSync gate test (normal pressure skips inside the cooldown, hard pressure past the backoff reaches the model)
• go test -race ./internal/agent/context/compaction/; go test ./internal/apperror/ ./internal/agent/runtime/native/ ./internal/handlers/ -count=1; pnpm --filter @memohai/web exec vitest run src/utils/api-error.test.ts (51/51) — all re-run on the combined head
• Full go test ./... + lint via pre-commit on each commit; two adversarial review rounds on the backoff (the original cap-equals-expiry sawtooth was found in review and is pinned by the rewritten cap test) and one on the error copy

⚠️ No human QA — this PR has not been verified by a human yet. Remove this line once a human confirms the happy path.