feat(report): show failed sweep cells with provenance

[dutifulbob]

Opened on behalf of Onur Solmaz (osolmaz). Work is ready for review.

Summary

Benchmark sweeps can have valid partial failures, especially when high-concurrency or high-context points hit memory limits.
This change keeps mixed-success sweeps reportable instead of treating one failed point as a failed overall run.
It also makes failed and skipped cells visible in the main HTML grid, and lets a reviewer click a cell to see the exact model/profile/workload settings and commands behind that result.

What Changed

The runner now distinguishes partial workload failures from all-failed runs.
A sweep with at least one completed measurement can finish as completed while still preserving failed/skipped measurements in the artifact.

• Changed final run error handling so partial sweep failures do not force a nonzero run error.
• Changed stored run status so mixed completed/failed sweeps are marked completed, while all-failed sweeps remain failed.
• Added failed/skipped labels such as mem floor, not ready, and server exit to headline report cells.
• Added native HTML detail disclosures for report cells with run, model, profile, workload, shape, profile config, vLLM serve command, benchmark command, engine args, serve JSON, and env JSON.
• Removed the command loading cap so larger sweeps keep command provenance available for every measurement.

Testing

I tested the changed runner behavior, the HTML rendering path, and the required local quality gates.
The dry benchmark path was also checked without starting a real model server.

• go test ./internal/report ./internal/vllmbench
• go test ./...
• go vet ./...
• npx -y @simpledoc/simpledoc check
• go run github.com/dutifuldev/slophammer/go/cmd/slophammer-go@v0.4.0 check .
• go run ./cmd/localperf bench run --dry-run --spec examples/diffusiongemma-vllm-standard/spec.json --profile 4k-reference --workload claim-repro-1k-out1024 --concurrency 1 --run-dir /tmp/localperf-onecase-dry
• go run ./cmd/localperf artifact check /tmp/localperf-onecase-dry.sqlite
• go run ./cmd/localperf artifact render /tmp/localperf-onecase-dry.sqlite --output /tmp/localperf-onecase-dry.html

Risks

The main behavior change is intentional: a run with partial failed points no longer exits as failed if at least one point completed.
All-failed runs and infrastructure errors still fail, so completely broken sweeps should not look successful.

The HTML report now includes more command/config provenance in hidden cell details.
That can make very large reports somewhat larger, but it is needed for clickable per-cell auditability.

[dutifulbob]

Final report:

• Implemented failed/skipped benchmark cells in the HTML report with native click-through details for status, reason, profile config, vLLM serve command, benchmark command, engine args, serve JSON, and env JSON.
• Partial benchmark-point failures stay reportable; all-failed sweeps and fatal run errors still fail the run. Fatal partial runs persist as failed.
• Aggregate repeat/model-level rows now say they are aggregates and no longer claim first-repeat run or command provenance.
• Dry-run planned rows stay neutral instead of rendering as failure labels.
• Secret-bearing command argv values are redacted before rendering in standalone HTML reports.
• Review loop: latest codex review --base main completed with no findings.
• Validation passed: go test ./...; go vet ./...; npx -y @simpledoc/simpledoc check; go run github.com/dutifuldev/slophammer/go/cmd/slophammer-go@v0.4.0 check .; one-case dry localperf bench run; localperf artifact check; localperf artifact render; dry-render grep for planned failure labels.
• CI: green (go checks passed).