Non-destructive apply + reaction-based scoring

[dsfaccini]

Two capabilities that make braindump usable incrementally against a live repo, rather than as a one-off dump.

1. apply — non-destructive merge into a repo

generate writes to staging (data/<owner>/<repo>/7-generate/). The new apply command merges those files into an actual repo without clobbering hand-written content:

uv run braindump --repo owner/repo apply --into /path/to/repo [--dry-run]

Every generated file is fenced between <!-- braindump: ... --> and <!-- /braindump -->. On merge, only the fenced block is replaced; content above/below is preserved. If the target has no markers the block is appended; if it doesn't exist it's created. Topic docs under agent_docs/ are now fenced too, so nothing braindump writes clobbers manual edits — this applies uniformly to every generated .md. Files braindump doesn't generate are never touched.

2. Reactions as a human verdict on review comments

A 👍/👎 on a review comment is an explicit human verdict — the highest-precision signal in the pipeline. download now fetches reactor attribution for any comment that has reactions (only reacted comments are queried, so it's cheap). It feeds two things:

• Reclaiming bot comments. Bot autoreview comments (github-actions[bot], coderabbitai[bot], …) are normally filtered out. A 👍 reclaims one as human-vetted signal, so you can curate a bot reviewer's output just by reacting to the good comments.
• Scoring. The net thumbs signal across a rule's source comments adjusts its score: endorsed 1.25×, vetoed 0.4×. An endorsed single-PR rule can clear the score floor it would otherwise miss; a vetoed rule drops below it.

--reaction-authors alice,bob restricts which logins count (default: anyone). Reactions are captured at download but the allowlist is applied at extract, so changing it only needs a re-run from extract.

Notes

• Score propagation: the reaction factor folds into calculate_rule_score, recomputed on dedupe merges (combining reaction_net across merged rules); place/group/generate inherit it with no changes.
• Backward-compatible with existing staged data — new fields default to 0/empty when absent.
• No test suite in this repo (per README), so changes were verified via targeted behavioral checks of the merge, reclaim, and scoring logic. Deferred (own follow-up): per-rule preservation via the <!-- rule:ID --> markers — block-level merge already protects hand-written content; edits inside the fence are still overwritten.