* First pass at recent_views stored in user local setting

This modifies the view log event handler to write a user's views to their local settings map (stored as json in the
appdb).

some notes on what causes a view

:card-query
 - occurs when collection with pinned cards/models is opened. maybe shouldn't count as a view

:card-read
 - happens when you view a card, like, load it in the UI and are looking at it

:card-read AND :card-query
 - happens when you view a model, like, load it in the UI and are looking at it

:dashboard-read
 - happens when you view a dashboard

:card-create
 - when you save a card
 - this will probably cause double counts because it will save and then trigger a :card-read right away too

:table-read
 - when viewing a table.
 - when selecting a db/table in dropdowns in the Question editor

I think we need the view log topics to be much smaller: #{:card-read :dashboard-read :table-read}

* WIP test for the recent_views user local setting.

This isn't totally correct yet. I think it doesn't clean up after itself correctly. That's pending.

And, I have to investigate because :table-read isn't working either. Still worth pushing this progress!

* Test that works! All potential view events need :actor_id

* Working test for /api/recent_views

* Order of views in user-local recent-views works now

If a user views something they've recently viewed, the homepage order doesn't change. Now it will, as in, the most
recently viewed item will always be the first item in the vector.

* One time recent_views expensive query done in the user-recent-views getter

The logic of filling the recent-views setting was first written in the activity api namespace. Now it's handled where
the setting is defined, which lets anyone else consuming the setting not worry about filling it if its empty.

The setting can be reset with `nil` which will cause the query to run again. Otherwise, if the list of recent items is
an empty vector, that is considered valid and will not re-run the query.

Updating the setting per view is handled still in the view log handler, implemented in the private function `update-users-recent-views!`.

* Add `most-recently-viewed-dashboard` setting

The most recently viewed dashboard for that user, in the past 24 hours.

This is not meant to be a permanent solution to providing this bit of info for the user, rather, it's a 'stop gap'
while the audit tables are reworked.

* Add test for most recent dashboard setting

* Pass :context to the view log fns

This lets us filter out views triggered by pinned cards.

* Fix tests failing after adding :context to view metadata

* See if card-query can help us filter out pinned item views

* Tests should send :card-query events

* Since we use :card-query, we have to filter out dashboard context too

This should prevent views being recorded if someone just opens a dashboard.

* ViewLog Cleanup to see if it'll let other tests pass

* drop non-viz percy tests, add a workflow that assigns visual label which triggers percy tests runs

* remove push trigger in require-percy-test workflow

* fix typo

* driveby type improvement

* Fix not showing single or multiple values options

When a parameter is connected to a question with template tags

* Review: unit tests instead

* Fixes inlining of numeric params in native queries
* Updates namespaces to use sql.qp/with-driver-honey-sql-version

Inferencing for models was switched from an LLM to an embedding selector. Unfortunately, this did not do a very good job as too much context was not retained. This change goes back to using the LLM for model inferencing.

* Backend for allow disabling auto-apply filters in slow dashboards (#30069)

* Dumb BE change before proper BE code is implemented

* Add some backend tests for auto_apply_filters

- included in revisions
- included in serdes

fixup some tests that didn't expect it

* Fix BE for public dashboards

---------

Co-authored-by: dan sutton <dan@dpsutton.com>

* Prevent auto apply filters on slow dashboards (#29569)

* Able to toggle auto-apply dashboard filters

Make FormField be able to show title with new style

* Only apply filters when clicking apply button

* Add E2E test

* Fix dashboard double fetch cards onload with parameters query

* Remove puzzle driven development comments

* Move logic from HoC to redux 1/x

* Make apply filter button works on embeddings

* Fix x-ray dashboard parameters not working

* Fix unit tests

* Add E2E tests to public dashboard sharing

* Add signed embedding E2E tests

* Fix the apply button transition not working when clicked

Since there's another CSS applied when the button is hovered.

* Reviews: Fix x-ray dashboard actions

Remove the need of 3-states boolean in `setParameterValue` action for
x-ray dashboards. This fixes so that BE sends x-ray dashboards with
`auto_apply_filters: true`.

* Review: remove duplicate `aria-label`

* Review: standadize `draftParameterValues` default values

* Review: Fix `aria-label` to not change when filter values change

* Review: avoid passing `ReactNode` as a prop

Also fix transition not working and apply button shift when adding a
filter.

* Move selectors and actions to their own file

* Review: Don't animate (un)mount of FilterApplyButton

* Address: Clean up eslint ignore on tests

* Review: Only test 1 scenario in E2E

Since testing other scenarios are unnecessary since they're repeat the
same thing

* Review: Use E2E helper to clean up test

* Review: Reuse selector

* Review: Use e2e helpers to clean up the test

* Fix typo

* Ensure Apply button won't appear unnecessarily (#30255)

---------

Co-authored-by: dan sutton <dan@dpsutton.com>

* Add GitHub issues config

This removes the need for a separate "Questions and help" issue template!

* Convert "Bug report" into a form

One of the challenges of metabot is creating input suitable for prompts of different sizes. You can, with very large models and dbs, create prompts that are not runnable. This PR contains a function that reduces the size of table prompts (this should actually happen very rarely, but it is ugly when it does) until it finally gets to a reasonably sized prompt or just silently returns nothing (in this case you won't be able to prompt engineer the particular table -- use a model instead). It also moves memoization to the right place for tables to reduce workload and capture the right segments of data needed for prompt engineering.

* Add `OrderByClauseDisplayInfo` type

* Add display info types

* Fix column group functions

* Update `displayInfo` function typing

* Make order-by functions aware of the stage index

* Remove duplicated `OrderByDirection` type

* Add missing newline

* Make order-by functions aware of sort direction

* Add wrapper for `change_direction` function

* Use MLv2 for `StructuredQuery's` `hasSorts` method

* Add `QueryColumnPicker` component

* Use MLv2 in `SortStep` component

* Disable `OrderBy` class validation method

* Patch E2E tests

* Prefer "order by" over "sort"

* Remove not longer relevant tests

* Fix pivot table breakouts check

* Adjust for `fk_reference_name`

* Revert E2E test changes

* Fix column group functions usage

* Revert redundant changes

* Remove alias `sort` method

* Remove `canAddSort` method

* Remove `sortOptions` method

* Stop using `clearSort` in the notebook editor

* Remove `cleanSorts` method

* Return plain MBQL from query `sorts` method

* Remove `OrderBy` class

* Mark remaining order-bys methods as deprecated

* Remove `addSort` usage (one more left)

* Remove `SortWidget` component

* Rewrite comment

* Fix import

* move pre-update-check-sandbox-constraints to defenterprise

* fix failing test so that it runs with :sandboxes feture enabled

* Don't load metabase.query-processor-test.test-mlv2 until we load QP

* Add schema and tests for `:value`

* Don't skip tests for #29895

* Conversion logic for `:value`

* pMBQL normalization should handle `:effective-type`

* Options should validate `:base-type`, `:effective-type`, etc. if present

* Use `:effective-type` for type checking if present

The current state is a collection of hacks to preserve some legacy
behavior (eg. normalization sets a blank `widget-type` to `"category"`)
and it doesn't work properly right now.

An example of a currently-broken case:
- `SELECT * FROM Orders WHERE {{tag}}`
- Set it to a Field Filter and select People > LONGITUDE
  - Or some other field for which type we don't have a widget.
- It shows "None" as the widget type, and the widget isn't shown at the
  top.
- Save the question and reopen it.
- Now the widget is visible (and a bit broken) as a `"category"` type.

This is because on reading from the appdb it gets normalized.

This PR uses `"none"` properly for cases where there are no widget
options for the field type (eg. longitude), and updates all the usage
sites to properly handle `"none"` rather than missing/undefined.

Existing saved questions with a blank `widget-type` will continue to
"work" as they currently do.

It was causing issues with query validation when it appeared in
unexpected places (like as the options on an `[:aggregation 0]`
reference.)

* Destructure the stage index in `LimitStep`

* Fix stage index management in notebook

* Generate step test IDs when building a step

* MLv2: finish implementing temporal bucketing

* Sort namespaces

* Sort namespaces

* Fix clj-kondo analysis for `defclause`

* Fix Kondo config

* Fix typo

* wait for db delete request to resolve before closing modal

* more async shenanigans

* Include checklist task to migrate off h2

New checklist item is shown when not hosted and app-db is h2. A
divergence from others, this checklist is not shown if it is not
triggered. In other words, we don't want to show this task as completed
in the UI if they are hosted or using pg, we just want to omit the task
completely. This is because we do not know if they upgraded or just
started with pg/mysql or started in the cloud.

Did lots of cleanup to make this easier to grok and easier to
test. Ultimately we don't have to redef a bunch of stuff to test the
scenarios.

* Fix setup links for external urls

* Re-export

* Add api types

* Add tests

* Improve tests

* Improve tests

* Improve tests

* Improve tests

* Improve tests

* Simplify if

* Show checklist as completed when on pg/mysql and not hosted

was filtering this out as we can't (easily?) tell if an instance was
migrated from h2 -> postgres. So wasn't clear if showing a checked-off
task list item made sense. But it does. So we'll show it. We just wont'
show it in cloud (hosted)

---------

Co-authored-by: Alexander Polyankin <alexander.polyankin@metabase.com>

Added embeddings to preprocess sql completions so that we downselect to an appropriate prompt size. This also gives us substantial performance improvements for model and table matching inference.

This reverts commit 90aa81de.