Per-language hook table for captureTagsFacts (#354 step 5) #374

Claim it in this issue first. Size: M. Depends on: nothing. Claimed by: nobody. Language: C++ (the work is in shared ingest code; one language per commit).

The problem, in user terms

Every language that is added to ripwire pushes a few more Lang:: branches into one shared function, so each new language edit touches code that every other language's answers depend on, and review has to re-check them all. captureTagsFacts (src/ingest_sidecap.h, about 815 lines, from void captureTagsFacts( to its closing brace) holds about 35 lines naming a Lang:: at c20f4d9b, Elixir the most (16 mentions), then Ruby, C++ and Clojure (4 each). Check the count on your checkout:

sed -n '/void captureTagsFacts(/,/^}/p' src/ingest_sidecap.h | grep -c 'Lang::[A-Z]'

Files to touch

• src/ingest_sidecap.h: captureTagsFacts. The decisions to move are per-capture: keep or drop the capture, which SymKind it gets, parameter count, arityExact, and the Ruby arms (bare-call capture, receiver written path, Rake/Jbuilder top-level self).
• a new src/ingest_<lang>.h per language that has a body to hold (Elixir's lives beside src/ingest_elixir.h), each hook with a one-sentence reason above it (CONTRIBUTING, "Hooks"). An empty slot means the shared path.
• the table itself: one constexpr hook table indexed by Lang, beside the other per-Lang tables (src/langfacts.h holds facts; this is hooks, so keep it out of that table). A value that is only data (a boolean, an enum) goes into langfacts.h as a column instead, per the "To decide where a change goes" rules in CONTRIBUTING.
• test/langidentity/ratchet.tsv: the ceiling falls with each removed predicate and may never rise; lower it in the same commit.

Red-first fixture and gate arm

This is a refactor, so the first artifact is the byte-identity harness, not a failing test: build main, then run bash test/langidentitycheck.sh BASE HEAD --langs=<the language you moved> --repo=<lang>:<your clone> after each commit. An empty diff is the pass. The red-first piece is a ratchet arm: the Lang:: line count in the function is asserted in a gate (a count that must fall), so a commit that adds a branch back goes red.

• Per-language gates that read the moved decisions: test/elixircheck.sh, test/elixirnamearitycheck.sh, test/elixirsemanticcheck.sh, test/clojurecheck.sh, test/rubybarecallcheck.sh, test/rubyrakejbuildercheck.sh, test/rubyrecvcheck.sh, test/kotlincheck.sh, test/jslangcheck.sh, test/langcheck.sh.

Acceptance

• Every language's output byte-identical (the langidentity run for the moved language, plus a before/after map and ripwire src --clones diff).
• One language per commit; Elixir first (it has the most).
• Lang:: lines inside captureTagsFacts fall to zero by the end; no new Lang:: comparison in a shared file.
• ripwire . --quality-delta=origin/main..HEAD gates 0; CONTRIBUTING "Definition of done for a common-code slice".

Not in scope

Moving captures into tags.scm (slices 4b and 5a are separate); the per-Lang fact tables (slice 6c); the base-clause readers.

Order with other work: src/ingest_sidecap.h is also touched by the open tags-predicate PR (#426, @conka8); rebase onto main before you start and again before review.