perf(nba): load roster once server-side, stop shipping full CSV to client

[FarhanAliRaza]

What

Refactors the nba template's backend to stop reloading and re-serializing
the entire player roster on every page load.

• Load the roster once at import time into a module-level PLAYERS
  constant instead of running State.load_entries on every page load and
  storing the full list in per-connection state.
• Stop sending the entire roster over the websocket on load. Because the
  full player list lived in state, all ~480 rows of CSV data were serialized
  to the client on initial connection. Only the current page (12 rows via
  get_current_page) now crosses the wire.
• Split the monolithic State into a table-only State and a chart-only
  StatsState, so chart data is never computed or sent to the client while
  the Table tab is showing.
• Replace pandas / numpy (used only for CSV parsing and np.random.choice)
  with stdlib csv and random, removing both dependencies.
• Collapse the duplicated per-chart aggregation loops into shared
  _passes_filters / _average_by helpers.

Impact (measured, prod build, 3 profiling passes)

Metric	Before	After	Change
WS payload onload	156.6 KB	17.0 KB	−89%

[greptile-apps]

Greptile Summary

This PR refactors the NBA template to load the player roster once at import time into a module-level PLAYERS constant, eliminating a per-connection CSV re-read and stopping the full roster (~480 rows) from being serialised over the websocket on every page load. The monolithic State is split into a table-only State and a chart-only StatsState, and pandas/numpy are replaced with stdlib csv and random.

• PLAYERS is a server-side constant; only the 12-row current page crosses the wire via the _filtered_sorted_players backend var, giving the 11x load-time improvement reported in the PR.
• Duplicated per-chart aggregation loops are collapsed into _passes_filters / _average_by helpers, substantially reducing backend.py.
• total_pages still computes from the fixed total_items = len(PLAYERS) rather than the current filtered count, so pagination controls over-report available pages during a search.

Confidence Score: 3/5

Safe to merge for the performance wins; the pagination count mismatch during search needs attention before it would be considered fully correct.

The roster load and chart-data refactoring are well-executed. The main concern is that total_pages derives from the static total_items rather than the length of the filtered player list, so typing a search term and then clicking Last Page navigates to an empty result set.

nba/nba/backend/backend.py — the total_pages computed var and the salary sort key for NaN sentinel players.

Important Files Changed

Filename	Overview
nba/nba/backend/backend.py	Core refactor: replaces pandas/numpy with stdlib csv/random, loads the roster once into a module-level PLAYERS constant, splits State into table-only State and chart-only StatsState, and adds _passes_filters/_average_by helpers. total_pages still references the fixed total_items instead of the filtered count, breaking pagination during search.
nba/nba/components/item_badges.py	Mechanical rename of State to StatsState for all badge event handlers and cond expressions; no logic changes.
nba/nba/components/stats_selector.py	Mechanical rename of State to StatsState for all selector event handlers and reactive bindings; no logic changes.
nba/nba/nba.py	Removes the on_load=State.load_entries handler; data is now ready at import time so no page-load event is needed.
nba/nba/views/stats.py	Removes the local StatsState definition (moved to backend.py) and updates all chart data bindings from State.* to StatsState.*.

Comments Outside Diff (2)

1. nba/nba/backend/backend.py, line 117-121 (link)

   total_pages is computed from the fixed total_items = len(PLAYERS) and never reflects the current search-filtered count. When a user types a search term that narrows the roster to, say, 5 players, total_pages still computes ~40 pages, so next_page and last_page happily navigate into empty slices. Since _filtered_sorted_players is already a @rx.var(cache=True), its length is the natural source of truth for pagination.

2. nba/nba/backend/backend.py, line 86-92 (link)

   When sort_value is "salary", players whose salary is stored as the sentinel string "NaN" produce float("NaN") → nan. Python's sorted() key comparisons involving nan are undefined — the relative order of NaN players versus any other player is unspecified and can differ across CPython releases and sort inputs. Consider placing them consistently at the end.

_{Reviews (1): Last reviewed commit: "refactor: drop pandas/numpy from nba tem..." | Re-trigger Greptile}