fix(clusters): keep the chosen region when a second region is removed on a single-region tier

[dawsontoth]

⊙ Problem

On a tier that allows only one region (the free tier, which carries allowedRegionIds), removing the second region of a new cluster threw away the first region too: it snapped back to the default US region and its smallest latency (#1275). The free tier's own validation error on the second row ("You can only select one region with this performance tier!") is what asks the user to remove it, so the natural fix-up step is the one that loses their choice.

Reproduced on current stage in a real browser (Playwright against a local dev server with mocked API responses): pick Medium, choose Europe 45ms for region 1, add a region, switch to Free, remove region 2. Region 1 read US / 90ms, small distribution afterwards.

The cause is autoSelectRegionBasedOnAllowedRegionIds in ClusterForm.tsx. It corrects the only remaining region whenever that region is not one the tier allows, and it always corrected to US. Removing the second row is what drops the count to one and fires it. The triage suspect, autoPickLatencyDescription in RegionFormInputs.tsx, is not involved: rows are keyed by field.id and removing a row leaves the other row's region and latency untouched on tiers without a region allow-list (pinned by a test below).

❓ Your call: the requirement as filed asks that the first region "remains". It now keeps its region and moves only the latency when the tier does not offer the chosen latency (Europe 45ms becomes Europe 60ms on the free tier), because the alternative is leaving an invalid selection with an error under it. If you would rather the form never auto-correct after a removal and let the user fix the flagged latency by hand, that is a one-condition change in the same effect.

💡 Solution

The correction now prefers an allowed entry in the region the user already chose, and falls back to US, then the first allowed region, only when the tier offers nothing there.

🔧 Changes

• src/features/clusters/upsert/ClusterForm.tsx — the one-line change in autoSelectRegionBasedOnAllowedRegionIds: a find on the surviving entry's regionName runs before the existing US and first-allowed fallbacks. Nothing else in the effect, including when it runs, changed.
• src/features/clusters/DESIGN.md — records when the correction runs (including on removal) and the same-region-first rule.

PR #1780 (custom regions) rewrites this effect to work on regionId; it keeps the US-first selection, so it will need the same preference expressed as "an allowed id whose catalog entry has the current entry's name" when it rebases. The two do not otherwise touch the same lines.

✅ Verification

Route: new component test driving the real ClusterForm (react-hook-form, zod resolver, Radix selects, the real Remove button) in jsdom, plus a live browser repro.

• src/features/clusters/upsert/ClusterForm.test.tsx (new) — five cases, reading each row's value from the hidden native selects Radix renders: removing a row on an unrestricted tier leaves the survivor untouched; on the free tier the survivor keeps its region whether row 2 or row 1 was removed ([Create Cluster] Deleting the second region will also revert the first chosen region and latency #1275); an already-allowed survivor is untouched; and a survivor in a region the tier does not offer still falls back to US.
• Fails on base: with the fix reverted, the two [Create Cluster] Deleting the second region will also revert the first chosen region and latency #1275 cases go red (expected [ 'US / 90ms, small distribution' ] to deeply equal [ 'Europe / 60ms, small distribution' ]), the other three stay green.
• npx vitest run src/features/clusters/upsert/ClusterForm.test.tsx → 5 passed (exit 0). Full gate via the pre-commit hook: vitest 396 files / 3668 tests passed, oxlint clean, dprint check --staged clean, commitlint ok; npx tsc -b exit 0.
• Browser: the same Playwright scenario after the fix ends with region 1 reading Europe / 60ms, small distribution; the four unrestricted-tier removal scenarios were unchanged before and after.

Cross-model review: two rounds, codex (graded) and gemini both ran. Their findings were rejected on evidence: a crash if firstRegion is undefined (the effect only runs when exactly one region exists, and the Remove button only renders with two or more), and explicit import extensions (this repo's imports are extensionless throughout). The Cursor leg could not fetch (1Password SSH agent refused to sign) and the Harper domain adjudicator failed authentication, so the outside findings were triaged by hand rather than adjudicated.

Closes #1275

_{🤖 Generated by Anthropic Claude Code (Claude Opus 5.5); posted via @dawsontoth.}

🤖 Generated with Claude Code

Related PRs: #1780 overlaps (rewrites the same correction effect for region ids)
Complexity: easy

_{Review-Coverage: authored=claude; ran=codex,gemini; blocked=cursor-composer(no-receipt),domain(auth); declined=cursor-grok,cursor-kimi,cursor-muse; rounds=2; full=1 @ fe4bbe8}

_{Review-Attention: skim ~2m (raised: degraded review) @ fe4bbe8}

[gemini-code-assist]

Code Review

This pull request updates the region correction logic in ClusterForm to prioritize keeping the user's chosen region when correcting regions on a tier that limits allowed regions, falling back to 'US' or the first available region only if the chosen region is not offered. It also adds corresponding documentation in DESIGN.md and a comprehensive suite of unit tests in ClusterForm.test.tsx to verify this behavior. There are no review comments to address.

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	68.96%	10312 / 14953
🔵	Statements	69.1%	11001 / 15919
🔵	Functions	62.36%	2666 / 4275
🔵	Branches	63.96%	7861 / 12290

File Coverage

File	Stmts	Branches	Functions	Lines	Uncovered Lines
Changed Files
src/features/clusters/upsert/ClusterForm.tsx	59.61%	41.81%	54.54%	59.73%	107-116, 121-125, 137-141, 160-164, 208, 211-224, 228, 231, 234, 282, 310-319, 338-371, 375-440, 459-463, 467, 471

Generated in workflow #2086 for commit fe4bbe8 by the Vitest Coverage Report Action