perf(memory): defer to_lowercase in extract_facts until triggers fire

[jony376]

Summary

Defers the allocating to_lowercase() in extract_facts until an identity, preference, or relationship trigger is present. Common no-match utterances (weather, questions, chit-chat) now skip the lowercase allocation entirely while preserving byte-identical extraction output.

Fixes #607

Changes

• needs_extract_facts_lower + ASCII case-insensitive trigger gates over the existing marker sets; to_lowercase() is built only when a gate fires.
• Remember-only fast path: Remember … with no other triggers still works without allocating a lowercase buffer.
• Extended the regression corpus with mixed-case no-match cases; updated extract_bench commentary.

Real Behavior Proof

• I have built and run the affected code locally (or noted why I could not).
• I have verified the change end-to-end on Jetson hardware.
• I have NOT verified on Jetson hardware, and I explain the equivalent verification path or validation gap below.

Tested profile / hardware (check all that apply):

• jetson
• raspberry_pi
• portable_sbc
• laptop
• mac
• CI-only / docs-only
• Not run locally

What I ran

No local Rust toolchain on this Windows dev machine (cargo not on PATH). Pushed branch and relied on upstream CI (cargo fmt, cargo clippy -D warnings, cargo test) for correctness proof. The extract_facts_corpus_regression integration test guards byte-identical output across identity, preference, relationship, remember, and no-match inputs including mixed case.

What I observed

Awaiting CI on push. Prior PRs (#562, #575) on the same repo validated the CI-only path; this change is confined to extract_facts with the existing differential corpus extended.

Test plan

• cargo test -p genie-core extract_facts_corpus_regression
• cargo test -p genie-core extract_
• Optional Jetson: cargo test -p genie-core --release --ignored --nocapture bench_extract_facts and compare no-match per-call latency before/after

Notes for reviewers

The trigger gates mirror the existing lowercase substring checks from #495; only the order changes (gate before alloc, not after). Relationship still uses the "my " marker gate already present in extract_relationships.

[matedev01]

LGTM — behavior-preserving (ascii-CI gating = old lowercasing for all-ASCII triggers), extended corpus regression covers no-match case. 826 lib, BFCL 96%.