refactor(models): keep release dates in one table / 模型发布日期统一到单一数据表

[Oseltamivir]

Summary

model-architectures.ts carried a releaseDate on each entry, and the unmerged fleet-lifecycle-section branch adds a MODEL_RELEASE_DATES table of its own. They disagree, and the disagreement is user-visible: the architecture table has DeepSeek-V4-Pro releasing 2026-06-08, six weeks after its own first InferenceX sweep on 2026-04-25, and that is the value rendered in the "Released by DeepSeek on …" caption.

Two copies of one fact drift. This PR leaves one table, on master, before the second copy can arrive.

No dates change here, and nothing users see changes. All eight move verbatim into MODEL_RELEASE_DATES — the wrong ones included — so the corrections land in a follow-up as reviewable data changes with sources, rather than buried inside a refactor. ModelArchitecture loses the field; ModelArchitectureDiagram reads getModelReleaseDate(model) and renders exactly what it rendered before.

What keeps them from diverging again

Structurally there is no second table, so the tests guard the ways this can still go wrong. Each was verified by mutation — mutant applied, named test observed failing, then reverted:

Guard	Mutation that kills it
An architecture entry may not carry releaseDate	re-add the field to DeepSeek_R1
Keys are display names, not DB keys	'DeepSeek-V4-Pro' → dsv4
Keys match a Model value	same
Dates are real calendar days	'2026-01-27' → '2026-02-30'
No date in the future	'2026-01-27' → '2027-01-27'

The typo guards matter because a bad key never throws: the lookup returns null and every caller silently falls back to its own default, so the Fleet Lifecycle axis would just quietly anchor to the wrong date.

MiniMax-M3 has a developer but no sourced date, so its caption does not render at all. A test names it rather than leaving that invisible.

Verification

• packages/constants unit: 39/39
• model-architectures.test.ts: 66/66
• bun run lint, bun run fmt, bun run typecheck: clean
• Pre-existing failures on master, untouched by this PR (confirmed by stashing and re-running on a clean tree): api-route-catalog shared-source digest, overview-data AgentX frontier, useThroughputData agentic interpolation.

中文说明

模型发布日期原本由 model-architectures.ts 每条架构自带，而未合并的 fleet-lifecycle-section 分支又新增了一张 MODEL_RELEASE_DATES，两者互相矛盾：架构表将 DeepSeek-V4-Pro 记为 2026-06-08，比其首次 InferenceX 基准测试（2026-04-25）还晚六周，而这个错误值正是页面上 "Released by DeepSeek on …" 所显示的内容。

本 PR 在第二张表出现之前，先在 master 上完成合并，只保留一张表。本 PR 不改动任何日期，也不改变任何界面：八个日期原样迁入 MODEL_RELEASE_DATES（含错误值），修正工作交由后续 PR 以「附来源的数据变更」形式单独审阅，而非夹在重构中。ModelArchitecture 移除该字段，架构图改为调用 getModelReleaseDate(model)，渲染结果与此前完全一致。

由于结构上已不存在第二张表，测试转而守卫其余出错方式，并逐一用变异测试验证：架构条目不得再自带 releaseDate；键必须是显示名而非数据库键，且必须对应 Model 枚举值；日期必须是真实存在的日历日，且不得为未来。键名写错不会抛错，只会返回 null 并让调用方静默回退，因此这类守卫尤为必要。MiniMax-M3 有 developer 但无来源日期，其说明文字不会渲染，测试将其显式列出。

验证：constants 单测 39/39，model-architectures.test.ts 66/66，lint/fmt/typecheck 均通过。另有三项单测在 master 上即为失败，与本 PR 无关（已通过 stash 后在干净工作区复现确认）。

---

Note

Low Risk
Structural refactor with no API or user-visible date changes; risk is mainly wrong keys silently returning null, which the new tests target.

Overview
Consolidates model release dates into a single canonical table in @semianalysisai/inferencex-constants (MODEL_RELEASE_DATES + getModelReleaseDate), instead of per-entry releaseDate on MODEL_ARCHITECTURES and a duplicate table on an unmerged branch.

ModelArchitecture drops releaseDate; ModelArchitectureDiagram now calls getModelReleaseDate(model) for the “Released by … on …” caption. Displayed dates are unchanged in this PR (eight entries moved verbatim, including known bad values slated for a follow-up).

Tests guard against re-divergence: architecture entries must not carry releaseDate; release-date keys must be display names matching Model enum values; dates must be valid calendar days and not in the future; MiniMax-M3 is explicitly listed as developer-without-sourced-date. api-route-catalog digest for models.ts is updated with a review note that no published API contract changed.

^{Reviewed by Cursor Bugbot for commit 5f0c09e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 15, 2026 6:03am

[Oseltamivir]

Rebased onto #742, which fixes the pre-existing master breakage that was failing this PR's checks. Base retargeted to fix/agentx-frontier-test-drift; please merge #742 first, then this retargets to master.

One addition while restacking: ../constants/src/models.ts is itself under the API-doc digest gate (review area "Published benchmark and TCO model names, aliases, and parameter enums"), and this PR edits that file. That failure had been masked because the overview-data.ts mismatch aborted the same test first. Reviewed: the change adds MODEL_RELEASE_DATES and getModelReleaseDate and touches no published model name, alias, or parameter enum — no endpoint exposes a release date — so the docs stand and only the digest moved. The reasoning sits next to it in api-route-catalog.ts.

Unit suite at this commit: 3367 app + 39 constants, all passing.

中文：已 rebase 到 #742（修复此前 master 上导致本 PR 检查失败的问题），base 暂改为 fix/agentx-frontier-test-drift，请先合并 #742。另外补充一处：../constants/src/models.ts 本身也在 API 文档摘要闸门内（审查范围为"已发布基准与 TCO 模型名称、别名和参数枚举"），而本 PR 修改了该文件；此前该失败被 overview-data.ts 的不匹配提前中断而掩盖。经复核：本次仅新增 MODEL_RELEASE_DATES 与 getModelReleaseDate，未触碰任何已发布的模型名称、别名或参数枚举，也没有任何端点暴露发布日期，故文档无需变更，仅更新摘要，理由已记录在 api-route-catalog.ts 中。当前提交单测：app 3367 项 + constants 39 项，全部通过。