Stop matching :bb in reader conditionals

[yogthos]

Fixes #893.

Stacked on #894 (base fix/wp-trace-getenv-load-order) — it re-records the same
two libconformance rows this touches, and one of them moves again here.

The reader feature set is {:jolt :clj :default} again. 0.7.10 added :bb on
the theory that a :bb branch solves the same non-JVM problems jolt has. It
does not. A :bb branch is written for babashka's host model, and the branch a
library writes for a host with no java.nio and no reflection is not the branch
for a host that has both.

Measured, matching :bb cost:

	with :bb	without
claxon (src)	FileNotFoundException: cheshire/core — will not load	loads
lasertag	58/49/9/0	58/51/7/0
tick	67/722/1/0	67/723/0/0
markdown-clj	94/180/0/0	94/182/0/0
cts (13 ns with :bb branches)	498 pass / 7 fail / 1 err	587 pass / 10 fail / 0 err
honeysql	177/2842/0/0	identical

• claxon and aws-api both write #?(:bb [cheshire.core] :clj [clojure.data.json]).
  jolt has no cheshire, so matching :bb turned a working library into a load
  failure.
• lasertag's two were :bb branches asserting sci.impl.fns class names and
  babashka's class-as-symbol hierarchies; tick's was a :bb branch asserting
  babashka's English-only locale rendering, where jolt renders the French
  through jolt-lang/time.
• On cts, descendants goes green — its 3 failures were the :bb branch.
  eq, not-eq and num pick up rows: three are the float/integer box model,
  and two are ##NaN divergences :bb had been hiding, now in
  known-divergences with :checks.
• honeysql orders #?(:bb … :clj (.sym ^Keyword k)) with :bb first at all
  three .sym sites, so jolt never reached its own keyword-direct-emit path.
  Same tally, and analyzer.clj/backend_scheme.clj no longer claim otherwise.

jolt.bb.fs and the loader's namespace-supplement seam existed only to fill the
babashka.fs/list-dir a :bb branch skips. babashka.fs defines it off its own
:clj branch now — the DirectoryStream import, Files/newDirectoryStream,
the reify DirectoryStream$Filter and (with-open [s …] (vec s)) all work
here, and fs-test passes unchanged. vendor/process goes back to upstream
babashka/process: the fork's only patch was a :jolt arm working around the
empty :bb splice, and process-test's timed-deref row pins the behaviour.

seqable? (first commit, independent)

seqable? was built out of coll?, so it answered false for a bare deftype
or reify declaring Seqable or Iterable — and for clojure.core.Eduction,
the one in core — while seq worked on all of them. malli's :every tests
seqability before it walks, so (m/validate [:every :int] (eduction …)) was
false. The predicate now reads the same per-method probes the seq arms read,
so the two answers cannot drift apart; jrec-declares-coll-iface? is
deliberately not it, since ILookup and Counted are in that list and neither
is Seqable on the JVM (corpus has the row).

:jolt/features

{:paths ["src"] :jolt/features [:bb]}

for a script ported from babashka whose :bb branches are the ones its author
wants. Additive — it can add a key, never remove one — and the project's alone,
like :jolt/replaces, because the feature set decides which branch every
library in the program is read through. Installed ahead of the class providers,
since those matter at the first class reference and this at the first form read.

malli

6757/36/34 -> 13261/28/10 — better on every counter. malli.sci's :bb branch swaps
m/eval's whole evaluator for load-string, so the suite's ~12k assertions
were running with m/eval stubbed and roughly half of them gated away
(ensure-parser-type is #?(:bb nil :default …)). Without :bb the suite
needs sci, and sci 0.12.51 from Maven does not load on jolt — its
copy-var splices the dereffed value of a macro var straight into the
emitted form, and a raw procedure is not a form jolt's analyzer takes. Filed as
jolt-l7tq; the recipe points at vendor/sci (the copy make sci and make scifunctional already gate), which clears all 72 ::sci-not-available errors,
and takes the timeout the full assertion set needs.

Three effects, no regression on any counter: half the suite was :bb-gated
(malli.parser-test alone goes 1754 -> 8224 passing assertions — its
ensure-parser-type helper is #?(:bb nil :default …), so every call was
vacuous), sci clears the 72 errors, and the seqable? fix clears the 104
identical? failures in parser-test plus 12 in core-test.

It is now the heaviest row in the gate by a wide margin: ~15GB RSS and ~30
CPU-minutes, hence :timeout 2400 and the runner's existing
JOLT_MAX_HEAP=off. Worth a look if CI machines are smaller than this one.

Gates

make test green — 104 ci targets + selfhost byte fixpoint. make libconformance on the seven affected rows: 3 ok, 4 better, 0 regressed.

Three things only the full gate caught, all in the last commit: the seed
records line numbers, so the comment edits in analyzer.clj and
backend_scheme.clj move it (make remint, converged in 2 passes); and
records-gambit.ss is generated from records-dispatch.ss among others, so
iface-seqable? had to be mirrored (make gambitgen, make gambitseed).

Site pages for cljc-interop, differences and tools-deps are in
jolt-lang/jolt-lang.github.io#29.

https://claude.ai/code/session_017f4uwd9Mq3Bn6NZLzBsjCT