fix(retro): persist recommendations so the next retro detects follow-through (#1834)

[jbetala7]

Problem

/retro generates its "3 Things to Improve" recommendations as prose output but never saves them to the JSON snapshot in .context/retros/. On the next run, Step 12 loads the prior snapshot for trend comparison but has no structured record of what it recommended.

The result: a week spent acting on retro feedback (security hardening, dependency bumps, the exact things the last retro asked for) gets mischaracterized as a generic "fix ratio is high" instead of "2 of 3 prior recommendations addressed." The improvement loop never closes.

Fixes #1834.

Root cause

The Step 13 snapshot schema (retro/SKILL.md.tmpl) persists metrics, authors, streak, greptile/backlog/test_health, and a tweetable — but not the recommendations. Step 12 reads the prior snapshot back for trend deltas, yet there is nothing in it to compare current-week work against.

Fix (repo-scoped /retro)

Three template edits in retro/SKILL.md.tmpl, then regenerated retro/SKILL.md:

1. Step 13 — the snapshot schema gains a mandatory recommendations array of {category, text} objects, populated from the "3 Things to Improve" items. Unlike greptile/backlog/test_health (omitted when empty), this is always written, because it is the input the next run reads back.
2. Step 12 — after loading the prior snapshot, each prior recommendation is scored against this window's commit subjects, changed files, and metrics as addressed / partial / open, surfaced in a Recommendation follow-through section ("2 of 3 prior recommendations addressed"). Snapshots written before this field existed have no recommendations key and are skipped silently, so old .context/retros/ history doesn't break.
3. "3 Things to Improve" narrative — wired to record the same 3 items into the snapshot array, closing the loop.

Matches the schema the issue proposed. Scoped to the repo-scoped retro that writes .context/retros/; global mode (~/.gstack/retros/) is left as a follow-up to keep this to one issue.

Testing

New test/regression-1834-retro-recommendations.test.ts, following the #1624 static-invariant convention — it asserts the full persist → read-back loop in both the template and the regenerated SKILL.md, so the capability can't silently regress:

bun test test/regression-1834-retro-recommendations.test.ts   # 8 pass

Verified the guard bites: 8/8 fail without the template fix, 8/8 pass with it. test/skill-validation.test.ts and the existing test/regression-1624-retro-stale-base.test.ts still pass.

Note: gen-skill-docs.test.ts has 8 pre-existing --host all / --host gbrain dry-run freshness failures on a clean main (c43c850c) on my machine, unrelated to this change — this PR adds zero new failures. The external-fork eval/report CI jobs may also fail on empty-env auth as usual; not actionable from a fork.

🤖 Generated with Claude Code

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here