fix(serenity): check retryOnQuota's shared deadline before every attempt

[byteclimber]

Summary

Self-review follow-up to #2874. Found that the shared per-request deadline was only checked inside the catch block, after a poll attempt had already failed — so the first poll attempt always ran regardless of how long ensure() itself took (e.g. queued behind other same-child recoveries on withResourceLock's own up-to-10s safety valve).

This undermines the deadline's whole purpose: a request where multiple stacked wrap sites (e.g. createProject + createPromptsByIds, both hit in the same generateTopics: true create-market call on a freshly-activated brand) each independently hit the disguised-405 lag could still compound latency past the intended budget — a narrower recurrence of the exact stacked-call-site risk the shared deadline was built to close (round-2 SRE review on #2874).

• Moves the deadline check to run before every attempt, including the first, not just between retries.
• Tracks the most recent metered-quota error (lastError) so a pre-attempt bail still throws something meaningful instead of needing a fresh failure to report.
• attempt now starts at 0 and increments only once an attempt is actually made, so a pre-attempt deadline bail reports exhausted at the attempt count actually reached (possibly 0, meaning ensure() alone consumed the whole budget).

Test plan

• npm run lint && npm run type-check && npm test — full suite green (15424 passing)
• Updated the two deadline-specific tests to reflect the corrected pre-attempt semantics
• Added a new test proving a budget already blown by ensure() skips the poll attempt entirely (the exact gap this fix closes)

🤖 Generated with Claude Code

[MysticatBot]

Hey @byteclimber,

⚠ Degraded review - no spec document was found for this change (searched the PR links, the touched repos' docs, the architecture/guidelines docs, and linked Jira). This review covers code-level quality but could not validate the change against an agreed design, so confidence is reduced. Add a spec link (PR template section 4) and re-request review for a full-confidence pass.

Verdict: Approve - clean, well-scoped fix that closes a real gap in deadline enforcement; no blocking findings.
Complexity: LOW - small single-service bug fix.
Changes: Moves the shared per-request deadline check to run before every poll attempt (including the first), so a slow ensure() call can no longer silently blow past the budget (2 files).

Non-blocking (2): minor issues and suggestions

• nit: The 8-line inline comment block (lines 181-192 in the new code) restates the design rationale already present in the header comment and includes process artifacts ("self-review finding", "found on self-review") that will age poorly. Consider trimming to 2-3 lines that state the invariant, not its discovery story - src/support/serenity/dynamic-allocation-active.js:181
• suggestion: When the pre-attempt deadline check fires at attempt=0, recordQuotaRetryOutcome('exhausted', { attempt: 0, callSite }) introduces a new metric dimension value that did not exist before (minimum was previously 1). Confirm downstream dashboards/alerts that consume this metric handle attempt=0 gracefully, or consider a distinct outcome label like 'budget_exceeded' - src/support/serenity/dynamic-allocation-active.js:190

---

_{Skill: pr-review | Model: us.anthropic.claude-opus-4-6-v1[1m] | Duration: 7m 59s | Cost: $3.35 | Commit: de25ef5a75da64003bc84d5ba950c513053e2589
If this code review was useful, please react with 👍. Otherwise, react with 👎.}