test: CI stays the exact-pin authority for Bun while local gates run one Bun per machine

[Blind-Striker]

What

The maintainer's rule is now one Bun per machine. Every machine runs Bun 1.4.3 in ~/.bun, while the pinned upstream (external/opencode/package.json) names bun@1.4.2.

Local gates therefore run on the machine's Bun through the existing OPENCODE_SDK_TESTS_BUN_MISMATCH=allow switch, and their results are local evidence only. CI stays the exact-pin authority. This PR adds a guard that keeps it so, and updates every piece of canon and every message that still told a host to put the pinned Bun first on PATH.

The change

• Guard: RepositoryGates/CiPinnedBunTests.
  • Ci_Should_Never_Set_The_Bun_Mismatch_Switch: no file under .github/ mentions the switch. The scanned set is asserted non-empty.
  • Ci_Should_Install_The_Bun_The_Pinned_Checkout_Names: setup-pinned-bun keeps bun-version-file: external/opencode/package.json.
• Canon.
  • quality-gates.md: "a host the maintainer has named" becomes the one-Bun-per-machine rule. The switch is set per command and never as a default, and its result is local evidence only. CI never sets it and is the exact-pin authority. The clone-gate paragraph says jscpd installs with the machine's Bun.
  • developing-on-windows.md: "the pinned Bun goes first on PATH" becomes "a workstation keeps one Bun and uses the switch".
• Messages.
  • PinnedServerPrerequisites: the version refusal no longer says "Put bun X first on PATH", which would mean installing a second Bun. It now names the switch. The missing-install refusal drops "with the pinned bun".
  • CloneDetector: its not-installed error and its summary say the machine's Bun, as does tools/jscpd/package.json's description.
  • A grep of tools/ and docs/ finds no "first on PATH" or "pinned Bun first" advice left.

Evidence that Bun 1.4.3 behaves like the pin

These runs were made on dev-01, which has only Bun 1.4.3, with the switch on, at RT-E2 C's head 656db84f for #442's investigation (#447):

• Launcher namespace on net10.0: 166 passed. The 4 failures were environmental: the service fixture was not built.
• Launcher namespace on net11.0: 170 of 170.
• All of OpenCode.Sdk.Tests on net11.0, run alone: 2798 of 2800. The 2 failures were environmental: the netstandard2.0 and net472 assets were not built.
• CI's exact parallel command, all modules: OpenCode.Sdk.Tests 2800 of 2800 on net10.0 and on net11.0. These are the pinned-server simulation and live rows.

No row behaved differently from the 1.4.2 runs.

Proof

• Red proofs for the guard rows. Appending the variable to ci.yml turns the first row red, and it names .github\workflows\ci.yml. Replacing bun-version-file with bun-version: 1.4.3 turns the second row red. Both files were restored.
• The new refusal assertion. FindAsync_Should_Name_Both_Versions_When_The_Bun_On_Path_Is_Not_The_Pinned_One gains "names the switch". Against the old message, PinnedServerPrerequisitesTests is 1 failed of 15; with the new message it is 15 of 15 on net10.0 and on net472.
• Checks. Whole-solution Release build: 0 warnings, 0 errors. The whole OpenCode.Sdk.Tools.Tests project: 2045 of 2045. verify-clones: 0 new. Slopwatch: 0. Format: clean.
• Local check: GATE PASS on eb004b16 (master 3f90dc7e): all 8 steps, and 10,514 tests with 0 failed.

Follow-ups

• Bun mismatch switch: drop it from local gates once the pin's packageManager names the machines' Bun #450: once the Faz 5 refresh pins the Bun the machines run, local gates stop setting the switch.
• The internal manual, the verify-array guidance and the dispatch policies switch to the machine's Bun, and the lead retires C:\wt\.tools\bun-1.4.2. That happens after this PR merges, outside this repository.

---

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