feat(api): retry failed and cancelled analysis runs on loopback

[seonghobae]

Why

GAP-003A still has no operator-visible retry path. docs/API_CONTRACT.md documents failed → retryable, and collection GET (#368) can list a failed or cancelled run, but POST /v1/analysis-runs with the original idempotency key returns the original receipt. Operators cannot start a new attempt without reconstructing the original body.

Live slices that this PR does not duplicate:

• feat(analysis): bind validation runs to scientific acceptance evidence #356 analysis_engine scientific-acceptance library
• feat(api): carry scientific acceptance only on terminal results #358 tepp_api terminal-wire DTO
• feat(api): serve scientific acceptance only on loopback status GET #359 GET loopback status by id
• feat(api): record scientific acceptance via loopback lifecycle POST #360 POST running/terminal lifecycle (stacked on feat(api): serve scientific acceptance only on loopback status GET #359)
• feat(api): cancel accepted and running analysis runs on loopback #361 cancel HTTP
• feat(api): drive scientific acceptance via loopback CLI #362 loopback CLI (ADR 0030)
• feat(analysis): bind TDT/CHRONOS composition to an analysis-run profile #364 GAP-007 TDT/CHRONOS analysis-run profile
• feat(api): enumerate analysis runs via loopback collection GET #368 collection GET (this PR stacks on that head 6a5d9777e2cf76c48726ae0243da872993d8881d)
• feat(persistence): persist idempotent analysis runs #287 GAP-003B persistence/Compose
• feat(network): replace union-find stand-in with Leiden consensus #351 Leiden

What

Eighth GAP-003A slice stacked on #368 over protected main 1bc02f580cf48e1d39da239f0e818453437c31c3:

• POST /v1/analysis-runs/{run_id}/retry on AnalysisRunLiveService
• clones a failed or cancelled run into a new metric-free 202 Accepted with a new run_id and new idempotency key
• parent remains failed/cancelled; collection GET lists both parent and child
• reusing the parent idempotency key fails closed
• accepted, running, succeeded, and unknown runs fail closed
• empty POST bodies admitted (path run_id + header new key)
• retry bodies and accepted receipts refuse RMSE/bias/coverage/SE-gate/scientific-acceptance/report keys
• consumer isolation; path/header/body identity mismatch fails closed
• GET-by-id remains 400 on this stack
• ADR 0032 (0026–0031 live on other GAP-003A PRs)

Not an ADR 0014 scientific claim. Not persistence. Not GET-by-id. Not lifecycle POST. Not cancel HTTP. Not collection GET. Not loopback CLI.

Verification

• cargo clippy -p tepp_api --all-targets -- -D warnings
• cargo test -p tepp_api --all-targets
• RUSTDOCFLAGS='-D warnings' cargo doc -p tepp_api --no-deps
• python3 scripts/validate_documentation.py
• python3 scripts/check_docstrings.py
• crates/tepp_api/tests/analysis_run_retry_http_contract.rs
• doctoring: docs/research/analysis-run-retry-http.md

Exact-head Checks on this SHA are required before merge. Two independent current-head APPROVEs are required. Author cannot self-approve. Merge #368 before this stacked PR, or retarget onto protected main after #368 lands. No force-push. Do not self-approve. Do not --admin merge.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}