Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@AI/backtests/compare_transformer_feature_schemas.py`:
- Around line 29-53: PRE307_DYNAMIC_23 lists MA feature names with underscores
(ma_5_ratio, ma_20_ratio, ma_60_ratio) but add_technical_indicators() actually
generates ma5_ratio, ma20_ratio, ma60_ratio so DataLoader.create_dataset
silently drops those features; update PRE307_DYNAMIC_23 to use the actual
generated names (ma5_ratio, ma20_ratio, ma60_ratio) OR alternatively change
add_technical_indicators() to emit the underscored variants so names match;
ensure the change is consistent with
transformer_refactor_change_analysis_2026-04-03.md and run the dataset creation
(DataLoader.create_dataset) to verify the three MA features are included.
- Around line 159-161: The dataset creation silently drops missing features;
update the compare logic after calling loader.create_dataset (the call that
returns X_ts, X_ticker, X_sector, y_class, _, info) to compute the difference
between requested_features and the actual effective_features returned in info
(or derive effective_features from what was used) and emit a warning log
including schema_name and the list of missing features; if all requested
features are missing, also surface that via a clearer exception or a stronger
log entry to aid debugging.

In `@AI/backtests/out/transformer_schema_compare_smoke/run_config.json`:
- Line 22: The run config currently hardcodes a local Windows user path in the
JSON key "output_dir", which breaks CI/other developers; update the "output_dir"
value to a relative project path (e.g., under ./out/) or make it configurable
via an environment variable and fallback to a relative path, and ensure the
generated smoke-test output directory is added to .gitignore so local artifacts
aren't committed; locate the "output_dir" key in run_config.json and replace the
absolute path accordingly.

In `@AI/libs/database/repository.py`:
- Around line 350-365: The else branch currently only deletes from executions
when account_id is set, omitting xai_reports, portfolio_positions, and
portfolio_summary; update the block around the cursor.execute calls that
reference account_id and target_date so it either (A) issues DELETEs for
xai_reports, portfolio_positions, and portfolio_summary using the same date
filter (without account_id) to perform the intended global-date removals when
allow_global_chain_reset=1 and account_id is provided, or (B) (preferred for
per-account isolation) add account_id columns to those tables and then update
the DELETE statements to include "AND account_id = %s" and pass account_id
(mirroring the executions delete). Modify the code paths that call
cursor.execute (the statements using account_id and target_date) accordingly so
the else block consistently handles those three tables.

---

Nitpick comments:
In `@AI/backtests/compare_transformer_feature_schemas.py`:
- Around line 97-100: The try/except around
tf.config.experimental.enable_op_determinism() currently swallows all errors;
update it to catch Exception as e and log a warning with the exception details
so failures to enable determinism are visible for debugging (use the module
logger via logging.getLogger(__name__) or the existing logger in this module,
and ensure logging is imported/available). This change should target the block
containing tf.config.experimental.enable_op_determinism to replace the silent
pass with a descriptive warning including the exception message/trace.
- Around line 255-261: Remove redundant int() casts for fields that are already
integers in the dict assignment (keys "n_samples", "lookback", "horizons",
"train_rows", "epochs_ran"): stop wrapping len(y_class), lookback, len(raw_df),
and len(history_df) with int(), and simplify the horizons construction by
removing the int(h) cast if horizons/h values are already ints (the code
building this dict — the lines assigning "n_samples", "lookback", "horizons",
"train_rows", "epochs_ran" — should return the raw integer values or an
identity-mapped horizons list instead of re-casting to int()).

In `@schema.sql`:
- Line 223: The column "account_id" is defined nullable—confirm this is
intentional for backward compatibility with legacy executions and ensure code
handles NULLs; update the schema definition around "account_id" to include a
clarifying SQL comment or inline note referencing account_names.id and the
compatibility rationale, and add a TODO/migration note to plan adding a NOT NULL
constraint once all executions are account-bound; also verify repository logic
that checks for "account_id IS NULL" (search for usages of account_id in
repository code) and adjust tests to cover both NULL and non-NULL scenarios.