feat(data): preserve additive append-only recipe variants

[cquil11]

Summary

• assemble a logical curve from consecutive same-image append runs plus the nearest full snapshot
• preserve each point's producer date, run, URL, and image while exposing separate logical curve identity fields
• add nullable recipe_fingerprint storage and identity so two additive recipes at the same topology and concurrency cannot overwrite each other
• propagate recipe identity through ingestion, latest/exact/history queries, APIs, overlays, charts, replay, and audited purge handling
• apply logical snapshots consistently to chart, Overview, calculator, exact-run, as-of, and historical consumers
• preserve legacy rows with a null fingerprint and keep fingerprint out of visual-series grouping

Why

Append-only producer runs contain only their new rows. Without logical snapshot assembly, the client treats that partial run as a new curve and drops inherited points. Without recipe identity, additive variants that share the App's prior coordinates can collide in ingestion or later consumers.

This PR separates:

• producer identity: where an individual point came from
• curve identity: which logical snapshot the point belongs to
• recipe identity: which generated benchmark recipe produced the point

Companion PRs

• InferenceX validation and sweep generation: feat(ci): support additive append-only curve updates InferenceX#2613
• disposable Qwen 3.5 canary: test(ci): validate append-only Qwen 3.5 AgentX point InferenceX#2614

Live preview

• filtered Qwen 3.5 AgentX curve
• new concurrency-6 point detail
• deployment: https://inferencemax-4wmqq1hrh-semianalysisai.vercel.app

Vercel team sign-in may be required because preview protection remains enabled.

The preview uses disposable Neon branch append-only-poc-20260814 (br-tiny-queen-aib16lin), which automatically expires at 2026-08-15 21:19:54 UTC.

End-to-end verification

• migrations 011 and 012 executed successfully against the disposable Neon branch
• GitHub sweep 31842466187 scheduled only B200 SGLang AgentX Qwen 3.5 TP4 concurrency 6
• measured profile duration: 303.07 seconds
• ingestion wrote one benchmark row and one append-only changelog entry
• the deployed exact-run API returns the concurrency-6 point plus inherited points from both producer runs
• the filtered curve page and point-detail page both return successfully
• a transactional seeded test also verified a new TP8 recipe and a same-config, same-concurrency fingerprint variant without collapsing either point

Test verification

• final focused database suite: 82/82 passed
• final focused App suite: 213/213 passed
• MCP suite: 25/25 passed
• formatting, lint, TypeScript, and git diff --check passed
• three separate independent reviewers completed extensive producer, App, and end-to-end audits; all actionable findings were implemented and the final verdicts are clean
• the full PR CI currently exposes pre-existing AgentX test expectation drift from merged PR Restore AgentX interactivity as the default metric #736; PR test: realign the AgentX frontier tests with #736 / 将 AgentX 前沿相关测试对齐 #736 #742 contains the exact test-only realignment and is green across unit and all browser shards

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 14, 2026 11:09pm

[blacksmith-sh]

Found 32 test failures on Blacksmith runners:

Failures

Test	View Logs
Inference CSV export with an unofficial-run overlay/ "before all" hook for "exports second-based latency headers and only currently visible overlay rows"	View Logs
Inference CSV export with an unofficial-run overlay/ "before all" hook for "exports second-based latency headers and only currently visible overlay rows"	View Logs
Official legend X works while an unofficial overlay is loaded/ "before all" hook for "shows official points and an official legend entry initially"	View Logs
Official legend X works while an unofficial overlay is loaded/ "before all" hook for "shows official points and an official legend entry initially"	View Logs
Overlay points follow canonical Optimal Only policy (agentic interactivity)/ "before all" hook for "hides trace-less overlay points in the default Optimal Only view "	View Logs
Overlay points follow canonical Optimal Only policy (agentic interactivity)/ "before all" hook for "hides trace-less overlay points in the default Optimal Only view "	View Logs
X-Axis Mode Toggle (inference chart)/ follows the percentile selector in the Interactivity axis label	View Logs
X-Axis Mode Toggle (inference chart)/ follows the percentile selector in the Interactivity axis label	View Logs
X-Axis Mode Toggle (inference chart)/ keeps the Advanced selection when the remaining tab is focused	View Logs
X-Axis Mode Toggle (inference chart)/ keeps the Advanced selection when the remaining tab is focused	View Logs

...and 22 more test failures. View all on Blacksmith

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9f82414. Configure here.}