fix(purchases): #667 recurrent -- recommend implementing the Lambda timeout bump + SDK retry tightening NOW (P0)

[cristim]

Symptom

Third recurrence of #667. QA reproduced again:

Failed to approve: execution 9336a111-85c5-4b74-9991-367350de9bac could not be approved:
some purchases failed: [t4g.nano: purchase failed: failed to find offering:
failed to describe offerings: operation error EC2: DescribeReservedInstancesOfferings,
https response error StatusCode: 0, RequestID: , canceled, context deadline exceeded]

Same shape as the prior #667 reproductions. StatusCode: 0, RequestID: "" = the AWS SDK request was cancelled at the network layer; the 60s Lambda budget got blown.

Evidence pulled from prod

• Execution row: plan_id: "", cloud_account_id: None, rec has no details/engine/cloud_account_id (i.e. a stripped-by-old-frontend payload, NOT a fresh post-fix(frontend/api): preserve details + engine on purchase POST (closes #597) #600 rec)
• Status stranded at approved (the bug(purchases): approve strands execution in 'approved' (purchase never runs, no error) on interrupted sync execution #632 stranding pattern that fix(purchase): reap purchase_executions stuck in approved/running >10m (closes #678) #681's reaper handles after 10 min)
• No purchase[9336a111-...]: log lines anywhere in CloudWatch. PR fix(purchases): add execution-tagged logging to executeSinglePurchase (refs #667) #668's exec-tagged logging merged at 12:56 UTC but the Lambda was last deployed at 16:23 UTC; this execution evidently fired BEFORE the deploy, so the logging didn't capture it.

What we still don't have

The smoking gun for #667's root cause (Lambda budget vs SDK retry storm) is the elapsed-time on the failed PurchaseCommitment call. Without #668's logging firing on a fresh failure, we can't confirm the timing hypothesis from CloudWatch — but the wall-clock fits: SDK default retry config is 3 attempts × 30s per call = up to ~90s, overrunning the 60s Lambda budget.

What to ship now

Implement the 2-part fix from #667 (still in PROPOSED state there):

1. Bump lambda_timeout 60 -> 300s in terraform/environments/aws/github-{dev,staging,prod}.tfvars for the api Lambda. Trades a 5x cost ceiling for resilience; Lambda Function URL has no upstream ceiling that would also need raising.
2. Tighten AWS SDK retry config on every purchase-path client (EC2, RDS, ElastiCache, MemoryDB, OpenSearch, Redshift, SavingsPlans) — pass aws.WithRetryMaxAttempts(2) + per-request http.Client{Timeout: 15 * time.Second} so transient slow API calls fail fast (15s × 2 = 30s) and surface a retriable error to the user instead of stranding the execution in approved.

After both ship + redeploy, re-test against the same SKU (t4g.nano in us-east-1) and confirm the next-cycle CloudWatch trace shows purchase[<id>]: PurchaseCommitment succeeded in <ms> OR failed after <ms> (the latter now bounded by 30s).

Acceptance criteria

• lambda_timeout = 300 in all 3 env tfvars
• EC2/RDS/ElastiCache/MemoryDB/OpenSearch/Redshift/SavingsPlans AWS SDK clients constructed with aws.WithRetryMaxAttempts(2) + per-request HTTP timeout ≤ 15s
• Regression test (Go): slow-API mock returns within ≤30s wall clock instead of SDK default 90s+
• Recommendation collection path is NOT regressed (it needs the higher retry budget; verify by checking providers/aws/recommendations/)
• Live verification: redeploy + a real t4g.nano purchase attempt either succeeds OR fails with a purchase[<id>]: ... failed after <elapsed> log line in CloudWatch

Cross-references

• fix(purchases): EC2 RI purchase times out under transient AWS SDK retries; Lambda 60s budget blown (P0) #667 (the proposed fix that this is the implement-now driver for)
• fix(purchases): add execution-tagged logging to executeSinglePurchase (refs #667) #668 (the exec-tagged logging — needs a fresh failure post-deploy to validate)
• fix(frontend/api): preserve details + engine on purchase POST (closes #597) #600 (frontend strip — may need a separate bundle-cache audit since this rec still lacks details)
• bug(purchases): approve strands execution in 'approved' (purchase never runs, no error) on interrupted sync execution #632 (status stranding — fix(purchase): reap purchase_executions stuck in approved/running >10m (closes #678) #681 reaps after 10 min, but the underlying SDK timeout is what needs to stop happening)

Severity

P0 / critical. Third recurrence. Every Approve click on a real t4g.nano (or similar latency-sensitive AWS API path) is at risk; the only protection today is #681's 10-min reaper, which cleans up but doesn't prevent.