fix(feature-flags): correct SDK method names, rollout weight units and broken references

[allen-mithra]

Summary

Correctness fixes in the feature-flag skills, found while auditing a local install. Following these passages produces wrong or failing output today.

• update-rollout weight units (launchdarkly-flag-targeting/references/targeting-patterns.md). The batching section told agents to turn a flag on and set a rollout "in one call to toggle-flag and update-rollout", with rolloutWeights of 10000 / 90000. update-rollout takes human percentages: its MCP description says "Weights must sum to 100. Use human-friendly percentages (e.g., 80 for 80%)", which also matches this skill's own SKILL.md. An agent following the reference would be off by 1000×. The section now says raw semantic-patch batching belongs in an approval request's instructions, and that direct changes are separate calls.
• Python and Ruby typed evaluation methods (launchdarkly-flag-create/references/sdk-evaluation-patterns.md, launchdarkly-flag-cleanup/references/sdk-patterns.md, launchdarkly-flag-drift/references/sdk-default-patterns.md). The server SDKs expose only variation, variation_detail and migration_variation: Python LDClient, Ruby LDClient. flag-create writes evaluation code from this reference, so bool_variation / string_variation / int_variation / number_variation / json_variation failed at runtime with AttributeError / NoMethodError. In cleanup and drift they were search patterns for methods that can't appear in code.
• React useFlags() (launchdarkly-flag-cleanup/references/sdk-patterns.md). The hook returns the flag map, so const { flags } = useFlags() destructures a key that doesn't exist. It now matches the shape launchdarkly-flag-create/references/sdk-evaluation-patterns.md already uses.
• Broken relative link (launchdarkly-flag-discovery/references/removal-readiness-checklist.md). The link ../launchdarkly-flag-cleanup/SKILL.md from references/ resolved inside the discovery skill. It's now ../../, as sibling references already do.
• Orphaned table rows (launchdarkly-flag-targeting/SKILL.md). A paragraph inserted mid-table left the "Full rollout" and "Copy from staging" rows rendering as stray text. The paragraph now sits below the table. Also removed an "Or by index" example in targeting-patterns.md that showed a variationId, not an index.
• Unit label (launchdarkly-guarded-rollout/SKILL.md). With 1000 = 1%, the unit is thousandths of a percent; a basis point is 0.01%.
• Worked example (flag-release/SKILL.md). The example holds production "until 2026-09-01", but its report line said 2026-08-01.

metadata.version is patch-bumped on each touched skill, and skills.json is regenerated.

Testing

• Not applicable

• Manual (describe below)

• python3 scripts/validate_skills.py: validated 49 SKILL.md files.

• python3 scripts/generate_catalog.py --check: up to date.

• python3 -m unittest discover -s tests: 9 tests OK.

• The Promptfoo evals were not run.

Notes

• Two tool names these skills reference don't appear in the hosted LaunchDarkly MCP server's tool list as of today: get-environments (flag-drift SKILL.md) and get-code-references (flag-discovery removal checklist). I left them unchanged, since they may target a different server version.

🤖 Generated with Claude Code

---

Note

Overview
This PR corrects agent-facing feature-flag skill docs so following them no longer produces wrong rollouts, broken code, or bad links. Touched skills get a patch bump in metadata and skills.json.

SDK references now match real APIs: Python and Ruby docs drop non-existent typed helpers (bool_variation, etc.) in favor of variation / variation_detail; React cleanup patterns use const flags = useFlags() instead of destructuring a non-existent flags property.

Targeting guidance separates MCP update-rollout (human percentages 0–100) from raw semantic-patch rolloutWeights (thousandths in approval batches), fixes a markdown table broken by mid-table prose, and trims a misleading fallthrough example.

Smaller fixes: guarded-rollout weight unit wording, discovery → cleanup link path, and a flag-release example date aligned with the hold scenario (2026-09-01).

^{Reviewed by Cursor Bugbot for commit 449f951. Bugbot is set up for automated code reviews on this repo. Configure here.}