Ninja generation and the scheduler do quadratic lookups #187

What

Two quadratic paths in ebuild:

1. NinjaBackend._write_ninja (ebuild/build/ninja_backend.py) matched each executable depends entry by scanning every target, and rebuilt the cflags string for every source.
2. run_graph (ebuild/core/scheduler.py) ordered freed nodes with order.index, a linear scan per node.

Measured (author's numbers, macOS, Python 3.14)

• ninja large (600 targets, 24k sources): 379.7 ms -> 195.0 ms
• scheduler fan-out n=4000, jobs=4: 98.5 ms -> 53.0 ms

Proposed approach

Build the static-library lookup once per manifest (a count, not a set, so output stays byte-identical when two static libraries share a name), build the cflags line once per target, and use a position map in run_graph. Add tests for internal static-library linking, per-source cflags, and parallel dispatch order. Generated build.ninja and compile_commands.json must stay byte-identical to master.

Status

Fix proposed in #186 (same author). Performance, not correctness.