fix(api): add per-stage node-rejection breakdown to placement failures

[AdaAibaby]

Fixes #3553

Rebased onto the placement error taxonomy (placement/errors.go + orchestrator/placement_errors.go) that has since landed on main. That taxonomy already classifies placement failures (timeout / no-capacity / no-compatible-node / unsupported-feature / create-failed), maps each to a semantic error_code and HTTP status, and forwards a curated, safe client message. This PR fills the two observability gaps it left, without touching what the client sees.

What

1. FailedToPlaceSandboxError now says why the fleet had no candidate. Previously it only listed the requirements (machine, labels, features). sample() now counts how many nodes each filter stage discarded — excluded, not-accepting, cpu-incompatible, feature-incompatible, label-filtered — and the error appends that breakdown.
2. The no-eligible-node return path in placeSandbox now logs. It was the only failure return with no logger.L() call (the create-failed and ResourceExhausted paths already log), so a sandbox_no_compatible_node failure left no trace in structured logs. Added a logger.L().Error with sandbox/template/build IDs and the full breakdown.

Why this does not re-open the topology-exposure concern

placementAPIError returns a fixed "Failed to place sandbox: no compatible node for this template's requirements" for this class and never calls err.Error(). The rejection counts and CPU details live only in the internal Err field, the OTel span, and the new log line — visible to operators, never to callers.

Example log / internal error

no node available with required metadata: machine={...}, labels=[gpu]; rejected 38 node(s): 0 excluded, 0 not-accepting, 38 cpu-incompatible, 0 feature-incompatible, 0 label-filtered

Changes

• placement/placement_best_of_K.go — add nodeRejectionCounts, tally per stage in sample(), thread into FailedToPlaceSandboxError, append breakdown in Error()
• placement/placement.go — log at the no-eligible-node site
• placement/placement_best_of_K_test.go — cover the breakdown in Error(), add sample() rejection-count test and end-to-end chooseNode propagation test

Testing

go test ./internal/orchestrator/placement/... passes (39 tests). go vet clean. Client-facing TestPlacementAPIError still passes, confirming the breakdown does not leak into ClientMsg.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ffeca63463

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3a6ab5b69e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[leonmeijer]

This change is already incorporated on main by 063aa20, including typed and sanitized placement-failure reasons plus regression tests. Merging this branch would duplicate the current implementation, so this PR can be closed as superseded.