Intercept .js and .jsx — no new grammar needed

[HackPoint]

What

Reads of .js and .jsx are not intercepted, so a large JavaScript file is billed at full cost
even though Lumen already carries a grammar that can outline it. It is the cheapest item left on the
coverage backlog: extension plumbing, not new parsing.

Evidence

• Live metering ledger: 97,678 tokens across 20 built-in reads of .js files at or above the
  300-line threshold, over 17 distinct files, the largest 1,232 lines. 57 .js reads in total;
  most fall below the threshold and are out of scope.
• tree-sitter-typescript 0.23.2 is already a dependency of lumen-core. Its TypeScript grammar is
  a syntactic superset of JavaScript; its TSX grammar handles JSX.
• crates/lumen-core/queries/typescript-tags.scm already captures class_declaration,
  function_declaration and method_definition — the node kinds JavaScript parses to under that
  grammar.
• Lists to change: crates/lumen-core/src/coverage.rs:32 (SOURCE_EXTS) and :41
  (INTERCEPTED_SOURCE_EXTS), plus both copies of the hooks' own list —
  .claude/hooks/lumen_read_intercept.sh:118 and the INTERCEPT_TEMPLATE the installer writes,
  lumenator/src-tauri/src/setup.rs:1266. Leaving the second one out recreates the drift
  coverage.rs was written to end.
• Interception is retroactive in the ledger metric, and this is the part that bites.
  coverage::classify tests today's INTERCEPTED_SOURCE_EXTS against every historical row, so the
  moment js enters that list the 20 above-threshold .js reads already recorded reclassify from
  UncoveredKind to Optimizable — which is precisely what leaked counts.
  crates/lumen-mcp/tests/efficiency.rs:585 asserts leaked == 0 across all rows with no era
  boundary, so that assertion fails on any machine carrying this ledger until the check is
  reconciled; SIGNED_SINCE at :439 is the existing precedent for drawing such a boundary.
  crates/lumen-core/src/coverage.rs:181 likewise asserts a.js at 500 lines is UncoveredKind,
  and has to be rewritten in the same commit.
• The repository corpus holds no real .js file at or above 300 lines, so this must be validated
  with fixtures rather than the corpus.

Not free. The mts/cts precedent is supported by the tools but deliberately not intercepted,
because interception changes which reads get diverted; the 97,678 tokens only land if js/jsx
enter both lists, which accepts that. Both known parsing failure modes degrade rather than break: a
mis-parsed file yields ERROR nodes, which the outline builder skips and the ranked arm declines as
NoDefs; minified bundles are single-line and never reach the threshold. The ledger
reclassification above is not a degradation — it is a test that goes red by construction, and
deciding how to reconcile it is part of this change rather than a surprise to discover afterwards.

Acceptance criteria

• lumen_core::structure::detect_lang maps js, mjs, cjs to TypeScript and jsx to Tsx.
• lumen_core::ranked::TagLang::detect returns the same mapping for those extensions.
• js and jsx appear in SOURCE_EXTS, in INTERCEPTED_SOURCE_EXTS and in both hook lists;
  the existing superset test still passes, and coverage.rs:181 is updated to assert the new
  classification instead of the old one.
• A .js fixture with a class, a function and a method outlines all three names, with
  returned_tokens < full_tokens.
• A .jsx fixture with a component outlines its name, with returned_tokens < full_tokens.
• The leak check in crates/lumen-mcp/tests/efficiency.rs is reconciled deliberately —
  reads recorded before .js was covered are distinguished from reads after it — and only then
  does the efficiency test report leaked as 0.
• .js leaves the uncovered-kind rows because it moved to the intercepted side, not because
  past reads were relabelled as defects.
• No new entry is added to any Cargo.toml.

How to test

cd /path/to/lumen
cargo test --release -p lumen-core structure
cargo test --release -p lumen-core coverage
cargo test --release -p lumen-mcp --test efficiency -- --nocapture
cargo tree -p lumen-core | grep tree-sitter

All three test commands must pass with the change complete — fixtures included, and with
coverage.rs:181 and the leak check updated as part of it. Run against unmodified assertions on a
machine that carries the ledger above, the second and third commands fail by construction, so a red
run there is the expected starting point rather than a regression. The efficiency run prints its
table: after reconciliation leaked must read 0 and .js must be absent from the uncovered-kind
rows. cargo tree must list the same grammars as before.