fix: expand {trackerInstructions} in the scheduled reference-watch prompt (#3140)

[atomantic]

Summary

reference-watch reaches an agent by two paths, and only the on-commit trigger (triggerReferenceAnalysis in server/services/referenceRepos.js) expanded the v3 prompt's {trackerInstructions} block. The WEEKLY scheduled path ran its own replace chain in buildImprovementTaskDescription (server/services/cosTaskGenerator.js) that did not include that token — so every scheduled run handed the agent a prompt whose ## Where to record proposals section was the literal string {trackerInstructions}. With no instruction for where to file, the agent had to guess, and on a GitHub/GitLab/JIRA-tracker app it would fall back to editing PLAN.md — exactly what the v3 prompt was written to stop.

• resolveReferenceWatchBlock now also resolves the app's work tracker (resolveAppWorkTracker) and returns trackerInstructions (via the shared formatTrackerInstructions) plus the isFileTracker-derived worktreeChangesExpected. Both helpers are reused rather than restated, so the block table and the flag keep one source of truth.
• {trackerInstructions} is substituted before {appName}/{repoPath}, so the injected block's own inner placeholders get expanded — the same load-bearing ordering the on-commit path documents.
• The generated task's metadata carries workTracker + worktreeChangesExpected off that same resolved value. This is the scheduled half of idle-no-changes fails reference-watch runs whose work tracker isn't the filesystem #3102: a forge-tracker run that files issues and edits no application code leaves a clean worktree, which the TUI idle-complete gate otherwise records as an idle-no-changes failure. PR stop failing idle-out TUI runs whose work product isn't files (#3102) #3139 fixed that for the on-commit path only.
• Non-reference-watch task types are untouched — the early return keeps the tracker resolution and the metadata stamp behind the existing type gate.
• Doc comments in referenceRepos.js and workTracker.js updated: both now name both dispatch paths as consumers rather than only triggerReferenceAnalysis.

Test plan

• New server/services/cosTaskGenerator.referenceWatch.test.js (9 cases): no literal {trackerInstructions} survives; the github block lands with its inner {appName}/{repoPath} expanded; workTracker + worktreeChangesExpected are stamped per tracker (github/gitlab/jira → false, plan → true); a stale schedule-level worktreeChangesExpected override loses to the resolved tracker (the flag must not contradict the instructions the agent got); a non-reference-watch type resolves no tracker and stamps no metadata; the no-reviewable-commits skip still returns null.
• Verified the new suite genuinely fails without the fix (reverting the one .replace line fails 2 assertions on the literal token).
• server/services/cosTaskGenerator*.test.js, referenceRepos.test.js, workTracker.test.js, cos.test.js, taskSchedule.test.js, cosValidation.test.js — 325 passed.
• Full cd server && npx vitest run: 22693 passed, 8 failed. All 8 failures are pre-existing and reproduce identically on clean main — they assert hardcoded Claude model ids / a data-root path that the ambient session environment overrides, and are unrelated to this diff.

Closes #3140