Make "most improved" reliable via a precomputed player_improvement view

Instead of dropping the question, give the model a clean path. The failure was
that "most improved" forced a hand-written two-season self-join, which errored
on the free model and fell back unguarded to fringe leaders.

New player_improvement view: season-over-season deltas (ts_pct, efg_pct, and
per-36 pts/reb/ast) joining player_advanced to the prior season (key derived
from the season string), floored at gp >= 20 both seasons so noise jumps can't
top it. A plain view over the small rollup — no game-log scan, stays fresh with
player_advanced. "Most improved" is now ORDER BY pts_per36_delta DESC.

Verified read-only: surfaces real breakouts (Alexander-Walker 13.4->22.5/36,
Duren, Brooks, MPJ), not the prior fringe noise. Prompt + schema updated to
route improvement questions here and forbid the hand-rolled join; restored
"Who improved the most since last season?" to the hint bank.

Migration (apply to Neon before prod):
CREATE OR REPLACE VIEW player_improvement AS ... (see db/enrich.sql)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/agent.py

@@ -239,6 +239,17 @@ player_advanced — one row per player per season (all seasons). Precomputed, so
   Leaderboards MUST filter gp >= 30 (and pts_per36 >= 15 for an efficient
   SCORER) — unfiltered, the TS% leaders are low-minute rim-runners.
 
+player_improvement — one row per player per season: season-over-season change vs
+  the SAME player last season. player_name, season, gp, prev_gp, and for ts_pct,
+  efg_pct, pts_per36, reb_per36, ast_per36: the current value, prev_<x> (last
+  season), and <x>_delta (current - prev; POSITIVE = improved). Already floored
+  at gp >= 20 in both seasons. Use for "most improved / breakout / fell off /
+  declined": just ORDER BY the relevant *_delta (DESC for improved). Do NOT
+  hand-write a two-season self-join — query this instead. "Most improved
+  scorers" -> ORDER BY pts_per36_delta DESC; "biggest efficiency jump" ->
+  ts_pct_delta DESC. (Only the last ~5 seasons of deltas are dense; a player
+  needs both seasons on record to appear.)
+
 team_advanced — one row per team per season (all seasons). Pace-adjusted:
   team (3-letter), season, gp, wins, off_rtg, def_rtg (LOWER is better),
   net_rtg, pace (poss/game), efg_pct, tov_rate, oreb_rate, ft_rate. Use for
@@ -422,8 +433,8 @@ their team" — pick the metric that actually answers it: an EFFICIENCY/RATE \
 rest of the league, clutch vs overall), or the relevant split. Examples: "best \
 clutch scorers" -> v_clutch by ts_pct; "most efficient scorers" -> \
 player_advanced by ts_pct; "is the Thunder offense or defense" -> team_advanced \
-comparing off_rtg vs def_rtg league ranks; "most improved" -> player_advanced \
-this season minus last.
+comparing off_rtg vs def_rtg league ranks; "most improved" -> player_improvement \
+ORDER BY pts_per36_delta DESC (never hand-roll the two-season join).
 - MINIMUM-VOLUME GUARD — mandatory on EVERY rate/efficiency/ratio/per-game \
 leaderboard, or fringe players with a handful of attempts dominate (a 2-for-2 \
 night reads as the "best shooter"). Always add BOTH a games floor AND a \

app/static/index.html

@@ -400,6 +400,7 @@ const HINT_BANK = [
   "Rank every team by net rating",
   "Curry's scoring efficiency in wins vs losses",
   "Rank every team by defensive rating",
+  "Who improved the most since last season?",
   "Best scorers per 36 minutes",
 ];
 const HINTS_SHOWN = 6;

db/enrich.sql

@@ -96,6 +96,30 @@ WHERE season_type = 'Regular Season'
 GROUP BY player_id, season;
 CREATE UNIQUE INDEX IF NOT EXISTS idx_player_advanced ON player_advanced (player_id, season);
 
+-- Season-over-season improvement, precomputed so "most improved / breakout /
+-- declined" is a trivial ORDER BY a delta — not a fragile two-season self-join
+-- the model writes by hand (which errored and fell back unguarded). A plain
+-- VIEW over the small player_advanced table (self-join on the prior season, so
+-- it stays fresh when player_advanced is refreshed; no game-log scan). The
+-- prior-season key is derived from the season string: '2025-26' -> '2024-25'.
+-- Floored at gp >= 20 in BOTH seasons so noise jumps (a 5-game sample doubling)
+-- can't top the board; rank by any *_delta. Positive = improved.
+CREATE OR REPLACE VIEW player_improvement AS
+SELECT
+    a.player_id, a.player_name, a.season,
+    a.gp, b.gp AS prev_gp,
+    a.ts_pct,    b.ts_pct    AS prev_ts_pct,    round(a.ts_pct    - b.ts_pct,    3) AS ts_pct_delta,
+    a.efg_pct,   b.efg_pct   AS prev_efg_pct,   round(a.efg_pct   - b.efg_pct,   3) AS efg_pct_delta,
+    a.pts_per36, b.pts_per36 AS prev_pts_per36, round(a.pts_per36 - b.pts_per36, 1) AS pts_per36_delta,
+    a.reb_per36, b.reb_per36 AS prev_reb_per36, round(a.reb_per36 - b.reb_per36, 1) AS reb_per36_delta,
+    a.ast_per36, b.ast_per36 AS prev_ast_per36, round(a.ast_per36 - b.ast_per36, 1) AS ast_per36_delta
+FROM player_advanced a
+JOIN player_advanced b
+  ON b.player_id = a.player_id
+ AND b.season = (left(a.season, 4)::int - 1)::text || '-'
+              || lpad((left(a.season, 4)::int % 100)::text, 2, '0')
+WHERE a.gp >= 20 AND b.gp >= 20;
+
 -- Advanced per-team-season rollup: pace-adjusted ratings + four factors (and the
 -- "allowed" mirror), via the Oliver possession estimate. Pace-fair, so it answers
 -- "is X more offense or defense", "fastest team", and opponent-strength tiers far