(OPT) Reduce compiler bootstrap memory without weakening executable owner coverage #798

Compiler bootstrap now consumes enough RAM to interrupt focused application-efficiency work on the shared 18 GiB machine. Reduce its peak without dropping production behavior, selected tests, private full-old oracles, hostile cases, assertions, or provenance checks.

Measured evidence on Rust 1.97.1, one Cargo worker, nonincremental, opt=0/debug=0, codegen-units=256, private SSD target:

• Source 376d774: production library completed in 84.67 s, sampled group peak 5,933,465,600 bytes (5.53 GiB). A one-second rustc stack sample during this build showed LLVM code generation; this does not establish the location of the overall peak.
• Production-library Clippy at d99fed0 completed analysis in 48.14 s, peak 4,165,746,688 bytes (3.88 GiB), with one intentional-inline-cursor lint subsequently corrected.
• Source 376d774: complete library-test typechecking reached a sampled group peak of 5,209,014,272 bytes (4.85 GiB) before four fixture errors; these are corrected at 705fdb7. No current test executable or runtime qualification is established by that attempt.

Receipts: /Volumes/Semaprax/lane-prompts/large-final-lint-repair-single-6g-20261012 and /Volumes/Semaprax/lane-prompts/large-final-serial-headroom-20261012.

Acceptance:

1. Identify the measured phase and a supported build/profile or implementation boundary that lowers peak memory; compare fresh matched builds on identical sources/settings and report time plus peak. Do not infer a compiler phase from aggregate RSS alone.
2. Preserve all production exports, deterministic artifacts and owning diagnostics. Keep complete independent saved oracles and the exact selected-owner discovery/counts. Any test partition must retain cross-module authority/source audits and full executable coverage, with no easier replacement corpus.
3. Preserve one-worker/shared-host safeguards, private targets, scoped Clippy, actual pin producers and exact build provenance. Avoid repeated production codegen while collecting test-fixture errors where ordering alone suffices.
4. Report bootstrap/setup costs separately from live-agent application input/authored tokens and accepted-task billing; no token-advantage claim from compiler-build memory savings.