Just like regular indexations, fulltext indexations are subject to last-commit-wins.
Which means that it is possible to reach a state where the fulltext table is
persistently desynchronised from ZODB:
- start fulltext indexation activity on many documents (typically: 100)
- modify one of the documents being indexed
- start fulltext indexation activity caused by this edit, and assume indexation only
happens for this object
- commit the single-object indexation (because it is very fast to retrieve fulltext
data from just one document)
- commit the many-objects indexation later (because it is much slower to
retrieve 100 fulltext representations)

As a consequence, cod emust spawn one fulltext indexation activity per
document, each with the appropriate serialisation tag. Serialisation tag which
must not conflict with regular indexation, so use a fixed prefix.
As a consequence of having to spawn one activity per document, use a
grouping method to still index by batches to amortise transaction overhead.
Keep the same method_id as before for backward-compatibility (maybe
dependencies on this value exist, even though it is bad practice).
Rewrite SQLCatalog_deferFullTextIndexActivity so ot works as a grouping
method, simplifying it in the process:
- build parameter_dict with all entries, as we already know all needed keys
- None is not callable, so test "not None" in just one expression
- remove whitespace at end of line
- use GroupedMessage API

''.split(',') is [''] which evaluates to false ...
e80b37ff was still incorrect

This was not really backward compatible with old slapos versions not passing
zserver_address_list

To answer question in the comment:
> How to inspect inside another iframe?

use selectFrame (after waiting for the frame to be added in the DOM of course).

Reference of selenium commands: https://web.archive.org/web/20160807232621/http://release.seleniumhq.org/selenium-core/1.0.1/reference.html#selectFrame (I'm not sure it's still on the web)

Tests asserts that size have all languages available, so let's be
compatible and just set the default language to the current selected
one at site creation.

This was not really backward compatible when running with old slapos,
because default value was not string in this case.

This step was in reality only removing the "Access Content Information" permission for Anonymous on web_site_module, which was not coherent with erp5_web, and had for consequence to make developers change it by hand when wanting to use web_site_module

Extend test runners to use partition IP  and run functional tests behind an https frontend.

See nexedi/slapos!374 for the slapos part and more detailed explanations.

/reviewed-on nexedi/erp5!729

/reviewed-on nexedi/erp5!742

/reviewed-on nexedi/erp5!737

I made a mistake rebasing & applying the patch to master, so the test
added in
nexedi/erp5!726 (comment 65067) did not
reach master.

This reapplies this part.

Thanks Seb for noticing !

On the traditional editor (from portal_component), selected text was not
highlighted

This is something the webmaster will define.
Also, initialize available languages like this will select a random
default language, which the webmaster might not understand.

So that a declaration like:
context.getUrlFor({command: 'display_action', options: {jio_key: "document", page: "action_reference"}})
can return the URL to redirect to the given action or the document

/reviewed-on nexedi/erp5!735

As it is backward compatible, this option can be passed as environment
variable.

Also re-order all argument output to make it clear which comes from
SlapOS generated wrapper and which are passed by test node.

We now access zserver behind the backend apache on https:// and this
apache's certificate is not trusted

Otherwise running in debug mode on headless machines just display
cryptic error message.

Auto-generate cache_id, removing the risk of duplicating one.
Inline getInstanceID as it is its only caller left.
Avoid wrapping if getTransactionalVariable cannot be imported.
Avoid intermediate variables, at the cost of dupicating logic but
accelerating run-time.
Use functools.wraps .
Access _cache_sequence_number directly - these pieces of code are friends
anyway, we already set a volatile property on Catalog instance. Saves one
getattr and one call.

No backslash at EOL.
Avoid splitting messages, to ease code lookup.
LOG's "error" argument accepts a boolean, which removes the need to touch
sys.exc_info.
Use modern exception raising syntax.

_getSelectDict was also calling itself on sub-sql_expressions, but this
was useless: they would already have called it, and discarded one of its
returned values. Instead, preserve both returned values in __init__ and
access these attributes directly when iterating over sub-sql_expressions.
Also, fold canMergeSelectDict back into its only 2 call places, saving one
getattr and one call.
Also, get rid of a few properties which were only used in _getSelectDict.
Also, get rid of two iterable copies: sql_expression_list when it is
already a list (no need to cast it again to a list, we just built one with
the list comprehension), and select_dict being cast into a list when
entering _getSelectDict, despite not being used (besides in extra
_getSelectDict calls, which are now entirely avoided).

Because some tests or softwares can chmod'ed some files to make them
readonly, testnode should try to chmod the files to deletable state if
deleting them fail in the first place.

Telephone.asUrl (which is used in erp5_short_message) properly supported
detailed coordinates, but had no support for coordinate text.

In that case, we don't try to be to clever and just use the coordinate
(almost) as is, just using the tel: prefix and removing spaces which are
not legal.

A pprofile run shows DateTime_or_None costing 3ms per value (including all
callees).
While this may seem low, it is common to receive hundreds of such values
in single query results (ex: web mode document & webpage lookups),
so the total time spent parting dates per query can become significantly
larger than the time to execute the query itself.
In the case of this measure setup, 504 dates were processed for a total of
1.5 seconds.
DateTime is slow to parse strings because it cannot rely on a pre-known
format. With MySQL/MariaDB, the format is know for those column types, so
implement specific parsing to get better performance.
On the same setup, pprofile run shows a parsing time of 123µs per value in
DATETIME_to_DateTime_or_None, or 65ms total.
DATE_to_DateTime_or_None should not deviate significantly from this.