[ohld]

Deletion-only cleanup of closed ranking experiments and flags that are off in production. The feed every user gets does not change.

What was deleted

Item	Evidence it was dead
recently_liked_blender_v2 (assignment, Redis 7d-LR quartile cache and lock, weights hook) in blender_experiments.py	Closed 2026-08-09, enrollment frozen. The weights hook already returned MATURE_BLEND_WEIGHTS for every user, which is what plan.blend_weights holds for mature users. The quartile cache had no other caller.
text_light_blender_v1 weights overlay and _fixed_pos_for_weights	meme_queue passed text_light_blender_v1_enabled=False. The fixed-position remap only fired when text_light_lr_smoothed was in the weights, which never happens on the growing or mature plans.
viral_shares engine plus the close-out stub get_mature_blend_weights_with_experiments	No blend weight referenced the engine. viral_shares_blender_v1 was killed on 2026-10-30.
cold_start_experiments.py and the fresh-viral branches in meme_queue / pipeline / candidates (max_age_days, the cold_start_explore_fresh* labels)	COLD_START_FRESH_VIRAL_EXPERIMENT_ENABLED defaults to false and is explicitly set to false in the prod env. The experiment was killed on 2026-10-30.
Source diversity (diversify_candidates_by_source), shadow score summary, full-payload sampler, Sentry span helpers, and their flags (RECOMMENDATION_SOURCE_DIVERSITY_ENABLED, RECOMMENDATION_SHADOW_SCORING_ENABLED, RECOMMENDATION_DIAGNOSTICS_SAMPLE_RATE)	Diversity defaulted to false and is not set in prod. The shadow score and full payload were diagnostics only and never changed ordering.
RECOMMENDATION_BLOCK_DISLIKED_* and block_disliked_sources_sql_filter (11 call sites)	Defaults to false and is not set in prod, so every call returned "".
Channel-viral broadcast pick and BROADCAST_CHANNEL_VIRAL_PICK_ENABLED	Off since 2026-10-30 (H3c) and not set in prod. The HQ pick is the default.
Tests: test_cold_start_fresh.py, test_block_disliked_sources.py, the v2/text-light cases in test_blender_experiments.py, test_meme_queue.py and test_pipeline.py, and the channel-viral cases in test_broadcast_pick.py	They covered the code removed above.

What stays

• explore_budget_v1 and everything it uses: assignment, fresh_viral engine and filter, bad-fresh drop, idle-return slot, labeling-budget low_sent query.
• The text_light_lr_smoothed engine, which is the cold-start CS3 weight and fallback.
• The disliked-source demote policy.
• The compact recommendation batch diagnostics log line. It still has the fields named in the explore-budget guardrail (selected_count, last_resort_used). It no longer has the keys for the removed features (source_diversity_*, shadow_*, cold_start_fresh_applied, the full_json payload).
• Analyst SQL and readouts for closed experiments. They read historical DB rows.

Verification

• Every removed symbol was grepped first. The flag values were checked in the production env.
• tests/recommendations, tests/feed_turn and tests/broadcasts, plus the sentry, experiment-assignment, redis, tgbot next/start/friend-challenge and crossposting tests: all pass locally (194 + 42 passed, 2 skipped).
• ruff check and ruff format are clean.

Diff stat

 .env.example                                       |   2 -
 specs/recommendations.md                           |   2 +-
 src/config.py                                      |  18 -
 src/recommendations/blender_experiments.py         | 445 +--------------------
 src/recommendations/broadcast_pick.py              |  86 +---
 src/recommendations/candidates.py                  | 124 +-----
 src/recommendations/cold_start_experiments.py      |  96 -----
 src/recommendations/meme_queue.py                  |  65 +--
 src/recommendations/pipeline.py                    | 440 ++------------------
 src/recommendations/utils.py                       |  46 ---
 tests/recommendations/test_blender_experiments.py  | 338 ----------------
 .../recommendations/test_block_disliked_sources.py | 140 -------
 tests/recommendations/test_broadcast_pick.py       |  72 ----
 tests/recommendations/test_cold_start_fresh.py     | 176 --------
 tests/recommendations/test_engine_contracts.py     |   1 -
 tests/recommendations/test_meme_queue.py           | 184 +--------
 tests/recommendations/test_pipeline.py             | 154 +------
 17 files changed, 70 insertions(+), 2319 deletions(-)

The COLD_START_FRESH_VIRAL_EXPERIMENT_ENABLED=false entry in the prod env can be removed after merge. It is harmless if left.