fix(aws): GetCapacityBlockOfferings/Reservations/Blocks drop per-region errors

[scttfrdmn]

Summary

• GetCapacityBlockOfferings buffered per-region errors, printed them only under --verbose, and always returned (results, nil) — the exact pattern already fixed in [HIGH] SearchInstanceTypes/GetSpotPricing return success even when every region query fails #63/Capacity Blocks for ML: instance-count gating, CB-only types, eligibility allowlist, 24h floor, utilization break-even #109 for other discovery paths, but this function was added afterward (Capacity Blocks for ML: wire the dead --blocks flag + add offerings discovery (DescribeCapacityBlockOfferings) #67) and reintroduced it.
• GetCapacityReservations and GetCapacityBlocks had the identical shape and were never fixed at all (investigation turned these up — the issue only reported the offerings path directly but asked to audit the others).
• Since "no capacity" is a common, legitimate answer for all three calls, a total failure (expired credentials, an SCP denial, throttling) was indistinguishable from genuinely empty inventory. All three now apply the same [HIGH] SearchInstanceTypes/GetSpotPricing return success even when every region query fails #63 contract as SearchInstanceTypes: an error when every queried region fails, an unconditional stderr warning for a partial failure.
• Adds aws.ErrCapacityBlockIneligible: DescribeCapacityBlockOfferings returns InvalidParameterValue with a distinct, stable message ("... is not supported for Capacity Blocks ...") when an instance type simply isn't a Capacity Block type at all — a different, permanent answer from a transient query failure, previously indistinguishable from one. Callers can check via errors.Is; when every queried region reports ineligibility (not a mix of ineligibility and real failures), the aggregate error names that specifically instead of the generic "all N region queries failed."

Verification against live AWS

Account 942542972736, us-west-2:

• c7i.4xlarge, m5.large, g7e.4xlarge, trn2.48xlarge → all return the exact InvalidParameterValue: ... is not supported for Capacity Blocks ... message the classifier matches.
• p5.48xlarge → hits a genuine CapacityBlockDescribeLimitExceeded service-limit error, correctly not classified as ineligible (verified isCapacityBlockIneligible returns false and the CLI surfaces the real AWS message unchanged).

Test plan

• TestGetCapacityReservations_Empty/_MultiRegion and TestGetCapacityBlocks_Empty asserted the bug — they only passed because Substrate doesn't implement DescribeCapacityReservations (answers InvalidAction) and the old code discarded that error. Replaced with _AllRegionsFailed/_MultiRegionAllFailed variants asserting the [HIGH] SearchInstanceTypes/GetSpotPricing return success even when every region query fails #63 contract, matching the precedent in TestGetCapacityBlockOfferings_AllRegionsFailed (Capacity Blocks for ML: instance-count gating, CB-only types, eligibility allowlist, 24h floor, utilization break-even #109).
• New pkg/aws/client_ineligible_test.go: classifier unit tests (ineligible vs. service-limit vs. access-denied vs. throttled vs. non-API error), allIneligible aggregation, and newUnreachableClient-backed all-regions-fail tests for GetCapacityReservations/GetCapacityBlocks.
• Confirmed regression: reverting just GetCapacityReservations/GetCapacityBlocks to the pre-fix implementation makes the new _AllRegionsFail tests fail with "expected an error ... got nil" — verified explicitly before restoring the fix.
• go build ./..., go vet ./..., go test ./... all green
• golangci-lint run ./pkg/aws/... — 0 issues
• make check-docs clean
• CHANGELOG.md updated under ## [Unreleased]

Fixes #110

[codecov]

Codecov Report

❌ Patch coverage is 68.18182% with 14 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/aws/client.go	68.18%	12 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!