fix(devx): give the driver-conformance census a dialect axis (#12134)

* fix(devx): give the driver-conformance census a dialect axis

The census scored `driver x case-set` and reported 45 covered cells before and
after a change whose entire content was "this conformance suite went from
executing on one dialect to executing on three". ADR-0053 D-A3 declares the
matrix with a third axis -- `driver {SQLite, Postgres at minimum}` -- enforced
only from inside a suite, by opt-in routing through live-dialect-matrix.testkit
and by OS_EXPECT_LIVE_DIALECT_MATRIX=1. Both fire only for a file that already
opted in, so a suite hard-coding `client: 'better-sqlite3'` was invisible to
every gate in the repo while counting as a covered cell.

Measured cost, not hypothetical: sql-driver-aggregation-conformance.test.ts was
green on main and on every PR; on live PG 16.13 sum/avg/min/max over a boolean
column threw SQLSTATE 42883.

DIALECTED is the new invariant. It asserts that a conformance suite SAYS which
dialects it runs on -- not that the answer is good enough, which is a different
decision. The population is defined by the presence of a DECLARATION, never by
the absence of a client literal, so moving the config into a helper cannot
respell a suite out of the gate's reach.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6HFzyH98W1YaQXhJUJt6o

* fix(devx): hold the dialect table's columns on a red run

The UNDECLARED glyph is four characters wider than the others, so the run where
the table is most worth reading was the one where its columns broke. Found by
the ablation, which is the only run that produces that glyph.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6HFzyH98W1YaQXhJUJt6o

* refactor(devx): read the dialect stance through the shared comment mask

The first cut of this axis carried its own stripComments. That is the exact
anti-pattern scripts/js-comment-mask.mjs was written to delete: every
source-scanning gate used to answer "comment or code" privately, and the copies
drifted into two silent families -- a naive regex that opens a phantom comment
on a `/*` inside a string, and a string-aware scanner that opens a phantom
string on a quote inside a regex character class. The shared module is validated
against @typescript-eslint/parser over the whole tree by
check-comment-mask-corpus.mjs in CI; a private copy is validated by whatever its
author thought of.

stripComments is the documented projection for this caller: it feeds a scanner
and reports neither a line number nor a byte offset.

The census output is byte-identical across the swap, so this is a change of
provenance and not of measurement.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W6HFzyH98W1YaQXhJUJt6o

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-warren