fix(ci): vendor autodoc prompt templates so docs regeneration works again

[anthony-at-pieces]

Summary

docs-freshness fails on main since the sampling-capability merge, and because the release workflow gates on CI success on main, releases cannot fire. The staleness itself is expected (lib/src/server and lib/src/types changed); the defect is that auto-autodoc's regeneration skips every module with Empty prompt for <type>, skipping, so the docs never regenerate and the hashes never advance. This PR makes regeneration work again by vendoring the prompt templates at the paths the config already declares.

Root Cause

.runtime_ci/autodoc.json declares its templates as consumer-repo paths (scripts/prompts/autodoc_quickstart_prompt.dart, ...) -- that is what the autodoc scaffolder writes -- but the scaffolder never writes the script files themselves. Resolution relied on a fallback in manage_cicd: when the relative path does not exist in the consumer repo, it resolves the runtime_ci_tooling package root from the consumer repo's .dart_tool/package_config.json and reads the scripts from lib/src/prompts/ there.

5e00fe5 removed runtime_ci_tooling from dev_dependencies, following the tooling's own v0.23.13 guidance ("use global activation only -- do not add as a dev_dependency"). Globally activated packages are not in the consumer repo's package_config.json, so the fallback returns null, templatePath stays scripts/prompts/..., and the prompt step runs dart run on a file that does not exist:

  [server] Pass 1: Generating QUICKSTART.md...
  [server] Empty prompt for quickstart, skipping
Generated docs for 0 modules, skipped 2 unchanged.
Autodoc failed for 3 module(s); hashes were not advanced

The sub-process failure is swallowed (runSync returns empty stdout), which is why the log says "Empty prompt" instead of "file not found". The last successful regen (00ef6ee, 2026-03-25 02:21 UTC) ran hours before 5e00fe5 (04:55 UTC) landed; every run since has this failure. The v0.23.13 guidance and the v0.23.13 template resolution contradict each other -- that tooling-side fix (resolve templates from the globally activated package, e.g. via its embedded-prompts extraction) belongs in runtime_ci_tooling.

Second layer, exposed once prompts generated: the workflows install the Gemini CLI unpinned (npm install -g @google/gemini-cli@latest). Current latest (0.56.x) rejects the --allowed-mcp-server-names= (empty) flag that runtime_ci_tooling's docsOnlyPolicyArgs passes for MCP isolation -- Invalid policy rule: mcpName is required if specified (cannot be empty) -- so every Gemini pass died at CLI startup (first commit's auto-autodoc run on this PR). The tooling's comment documents the old contract ("Gemini CLI 0.34+ behavior: empty list -> none"), which no longer holds.

Fix

Vendor the four prompt scripts from runtime_ci_tooling v0.23.13 lib/src/prompts/ -- _ci_config.dart, autodoc_quickstart_prompt.dart, autodoc_api_reference_prompt.dart, autodoc_examples_prompt.dart -- verbatim (formatted at 120 cols) into scripts/prompts/, the exact paths autodoc.json already declares. Template resolution now takes the primary consumer-repo branch (File('$repoRoot/$templatePath').existsSync()), before the broken fallback is ever consulted. The scripts are self-contained (dart:io, dart:convert, and the sibling _ci_config.dart only). scripts/** is excluded from analysis so the vendored files stay byte-diffable against upstream.

The source variants are vendored rather than the tooling's embedded (AOT) variants: the embedded quickstart prompt is trimmed (no changelog-context section), and the source variants are what produced the existing March docs.

Second commit: pin @google/gemini-cli@0.35.0 (the @latest of the last successful regen, 2026-03-25, and the version v0.23.13's flags were written against) at all 9 install sites across ci.yaml, issue-triage.yaml, and release.yaml, so one upstream CLI release cannot break every Gemini-backed job at once. These are template-generated files; the durable fix belongs in the templates (tracked with the resolver bug in open-runtime/runtime_ci_tooling#70).

Test plan

• dart run scripts/prompts/autodoc_quickstart_prompt.dart Types <repo>/lib/src/types <repo>/lib/src/types -- exit 0, 8904 bytes of prompt (previously: file did not exist, empty stdout)
• Same for autodoc_api_reference_prompt.dart -- exit 0, 1871 bytes
• dart pub global activate ... --git-ref v0.23.13 + manage_cicd autodoc locally -- reaches the Gemini gate ("Gemini CLI not installed") instead of failing at prompt generation; the Gemini half only runs in CI where the CLI and GEMINI_API_KEY exist
• dart analyze -- 71 issues, identical to main baseline (scripts/** excluded)
• dart format --line-length 120 --set-exit-if-changed on the vendored files -- 0 changes
• First commit's auto-autodoc run: prompt generation succeeded for all modules (Pass 1: Generating QUICKSTART.md... with no Empty prompt), proving the vendored templates resolve; Gemini then crashed at startup on the unpinned-CLI flag incompatibility above
• This PR's auto-autodoc run on the second commit is the end-to-end proof: regenerate server / types / top_level docs, push a bot(autodocs) commit to this branch, docs-freshness green (reviewer-visible in this PR's checks)
• Tooling bug filed: open-runtime/runtime_ci_tooling#70 (global-activation guidance vs package_config-based template resolution); once fixed there, this vendored copy can be dropped or kept as an explicit override

[anthony-at-pieces]

Executive summary, since the diff looks bigger than it is.

What actually changed (556 lines, ~10 of them decisions):

• 546 lines are 4 files copied verbatim from runtime_ci_tooling v0.23.13's own lib/src/prompts/ into scripts/prompts/. Zero original code -- review is "diff against upstream", not "read 546 lines".
• 9 one-line version pins: gemini-cli@latest -> @0.35.0 in the generated workflows.
• 1 analyzer-exclude line.

How it broke, and when:

1. On 2026-03-25, 5e00fe5 removed runtime_ci_tooling from dev_dependencies -- following that tool's own v0.23.13 upgrade guidance. Side effect nobody could see: the autodoc config points at prompt scripts under scripts/prompts/ that were never in this repo; the tool silently found them through the dev_dependency. With it gone, every autodoc run since has generated empty prompts and skipped all modules ("Empty prompt for quickstart, skipping"). Invisible while no source changed.
2. PR fix(mcp): model sampling tools and context as object capabilities #10 changed lib/src/server + lib/src/types, so the docs are now legitimately stale, regeneration is broken per (1), and docs-freshness fails on every main run.
3. Fixing (1) exposed (2): the workflows install gemini-cli@latest, and the current latest (0.56.x) crashes on a flag the March-era tooling passes. Pinning to 0.35.0 -- the version running when autodoc last worked -- fixes that.

Why fix it at all:

• main CI is permanently red, and the release workflow only fires when main CI is green -- so no automated releases until this lands.
• It self-heals the docs too: once merged, the next autodoc run regenerates the server/types docs for the fix(mcp): model sampling tools and context as object capabilities #10 changes and advances the hashes.

Blast radius: CI/docs tooling only. No lib/ code, no runtime behavior, nothing shipped to consumers. Both root causes are also filed upstream (open-runtime/runtime_ci_tooling#70) so the vendored copy and pins can eventually go away.

[anthony-at-pieces]

I'm open to different approached to fixing this, but currently I'm downgrading this in priority to focus on cost saving issues and will come back to this.