Deprecate SIMULATE_CHOOSER_COLUMNS and LOGSUM_CHOOSER_COLUMNS settings

[Copilot]

SIMULATE_CHOOSER_COLUMNS and LOGSUM_CHOOSER_COLUMNS were memory hacks that manually pruned chooser tables before merging with alternatives. util.drop_unused_columns (on by default via compute_settings.drop_unused_columns) now does this automatically from the spec, so the manual lists are redundant — and a well-known gotcha, since omitting a column used by a spec silently breaks the model.

Settings

• Both settings are now typed Any | None with a mode="before" field validator that emits a DeprecationWarning and discards the value. Existing configs continue to load rather than tripping extra="forbid".
• Applies to TourLocationComponentSettings, TourModeComponentSettings, TourSchedulingSettings, and SchoolEscortSettings.

Component code

• Removed the # FIXME - MEMORY HACK column subsetting (and the associated # Drop this when PR #1017 is merged household_id workarounds) from location_choice, util/tour_destination, util/tour_od, util/tour_scheduling, and school_escorting.
• Deleted the now-unreachable helpers logsums.filter_chooser_columns and expressions.filter_chooser_columns.

Configs / docs / tests

• Stripped both settings from the 46 bundled example config files.
• New activitysim/core/test/test_deprecated_chooser_columns.py covering the warning per settings class and the no-warning-when-omitted case; updated the two existing tests that supplied the setting.
• Change log entry under "Upcoming Changes".

>>> TourModeComponentSettings.model_validate(
...     {"SPEC": "tour_mode_choice.csv", "LOGSUM_CHOOSER_COLUMNS": ["age"]}
... ).LOGSUM_CHOOSER_COLUMNS
DeprecationWarning: LOGSUM_CHOOSER_COLUMNS is deprecated and no longer used,
unused columns are now dropped automatically
None

Users should remove these keys from their configs; no other action is required.

• Fixes Simulate Chooser Columns and Logsum Chooser Columns Settings Not needed #1016

[jpn--]

activitysim/abm/models/util/tour_destination.py:697  — The change merges all  persons_merged columns into tours instead of retaining the previous narrow chooser-column selection. During household tracing,  interaction_sample  disables automatic chooser-column pruning, then cross-joins these wide choosers with alternatives. Because  persons_merged  can include household, land-use, accessibility, and disaggregate-accessibility fields, destination sampling can regress to a large chooser×alternative memory footprint and potentially OOM realistic traced runs.

Retain a safe explicit chooser-column subset for trace/estimation paths, or change pruning to preserve only trace-required fields rather than bypassing pruning completely.

[jpn--]

@copilot, address review comments on this PR

[Copilot]

activitysim/abm/models/util/tour_destination.py:697  — The change merges all  persons_merged columns into tours instead of retaining the previous narrow chooser-column selection. During househ...

Fixed in 71f88ba by taking the second option: _interaction_sample no longer bypasses pruning when trace targets exist. It now always runs util.drop_unused_columns on both choosers and alternatives when compute_settings.drop_unused_columns is enabled, adding the trace slicer id columns (person_id, proto_person_id, proto_household_id; household_id is already always retained by drop_unused_columns) to the protected list so interaction_trace_rows can still identify traced rows.

The full, unpruned choosers and alternatives tables are still written to the trace output earlier in the function, so nothing is lost from the trace except extra unused columns in the interaction_df dump. Estimation paths are untouched — interaction_sample passes estimator=None to utility evaluation and the other call sites that guard on estimator is None were not changed.

[jpn--]

•  activitysim/abm/models/util/tour_destination.py:697  — The change merges all  persons_merged  columns into tours instead of retaining the previous narrow chooser-column selection. During household tracing,  interaction_sample  disables automatic chooser-column pruning, then cross-joins these wide choosers with alternatives. Because  persons_merged  can include household, land-use, accessibility, and disaggregate-accessibility fields, destination sampling can regress to a large chooser×alternative memory footprint and potentially OOM realistic traced runs.

Retain a safe explicit chooser-column subset for trace/estimation paths, or change pruning to preserve only trace-required fields rather than bypassing pruning completely.

[jpn--]

The current release is v1.5, so this will be officially deprecated as of the next release (1.6)

[jpn--]

The current release is v1.5, so this will be officially deprecated as of the next release (1.6)

[jpn--]

This looks good to me. The SANDAG test is failing due to the bug in at-work tour logsums.

[jpn--]

This looks like the AI (here, OpenAI's Sol) found and fixed a mostly unrelated bug.

[jpn--]

@dhensle, @janzill I don't think a comprehensive review from both of you here is totally necessary, but I know each of you has been in some of these files recently (and this is a re-do of David's original PR from some time ago) so I want to give you an opportunity to comment / review before I finalize and merge this.

[janzill]

this all looks good, no suggestions but one comment: Sol did well finding that at-work origin bug.

[dhensle]

This looks good to me too. Glad to see copilot kept the traceable id columns which I think has previously been an issue that was also fixed as part of this PR. This is not tested as part of this, but we should really add some better tracing test coverage. I think that is better done in a separate PR.