chore(graph-builder): remove five shadowed/orphaned methods (-552 lines, no behaviour change) - #1619

…P graph_name, JS/TS imports, .NET ignore defaults and dev-env pins (#1646)

* fix(jobs): stop cleanup_old_jobs deleting jobs that are still active (#1537)

Active jobs with recent progress survive the sweep; a RUNNING/PENDING job
with no update for the whole window is flipped to FAILED (still visible to
check_job_status) instead of vanishing, and ages out via the terminal path.

* fix(watcher): clear watched_paths and watches in stop() (#1519)

A stopped Observer cannot be restarted; stale entries made watch_directory
answer 'Already watching' for a dead observer.

* fix(validator): stop rejecting read queries that mention write keywords as identifiers (#1511)

Keywords in provably-safe positions (property access, label/rel-type,
parameter, AS-alias) no longer trip the read-only gate; clause-position
writes are still rejected, including a real SET/DELETE following an alias.

* fix(simulator): make simulate_metrics work on embedded backends (#1512)

labels(<var>)[0] is now rewritten for any variable name (the literal
labels(n)[0] replace missed n1/n2), and the relationship query's bare
pattern predicate is restructured as a CONTAINS*0.. match.

* feat(mcp): honour add_code_to_graph graph_name on multi-graph backends (#1558); drop last shadowed method (#1604)

GraphBuilder accepts a graph_name and binds its writer to
get_driver(graph_name); the handler builds a scoped builder on FalkorDB and
refuses explicitly on single-graph backends instead of silently indexing
into the default graph. Also removes the remaining shadowed
add_repository_to_graph definition #1619 missed.

* fix(ignore): exclude .NET obj/ from the default ignore patterns (#1585)

obj/ contains generated source (AssemblyInfo.cs, GlobalUsings.g.cs) that
added phantom File/Module/IMPORTS rows to the graph. The golden test already
masks obj/ via GENERATED_BUILD_DIRS, so the drift never failed the test —
it only rotted the recorded golden. With the indexer itself excluding obj/,
raw output matches the comparison scope and the C# golden regenerates with
zero logical drift (the CALLS 12→7 loss reported in #1585 no longer
reproduces on current main; it was fixed by intervening parser merges).

* fix(js/ts): extract named, default and namespace ES import bindings (#1526)

import_clause is an unnamed child of import_statement, so
child_by_field_name('import') always returned None and every ES import
collapsed into the bare-module fallback — the entire binding-extraction
block was dead code, and would have tested the wrong node types anyway.
The clause's children are now walked (a single clause can carry a default
and named bindings together), the namespace alias is read from the
identifier child, and both grammars are covered by binding-level tests.

Downstream effects, captured in the regenerated goldens: IMPORTS edges
carry the real binding in imported_name (was the module path), and 4
TypeScript call edges upgraded from HEURISTIC_CALLS to resolved CALLS
because named-import bindings now reach the imports map.

* test(env): fail loudly when tree-sitter-language-pack violates the pyproject pin (#1625)

A stale venv holding an out-of-pin grammar pack makes the parser suite pass
while disagreeing with CI (the #1600 misparse exists on 1.14.x but not
0.13.x, which nearly got a correct fix closed as a no-op). The divergence is
now a named test failure pointing at the venv, not the parser.