feat(handlers): xcodebuild noise filter + log-file tail detection

[claudioemmanuel]

Summary

• xcodebuild handler: BuildHandler now detects xcodebuild and strips 30+ Xcode-specific action prefixes (CompileSwift, SwiftEmitModule, Ld, CodeSign, ClangStatCache, build-system preamble, invocation echoes, etc.) while preserving Swift diagnostics and the terminal ** BUILD SUCCEEDED/FAILED ** markers.
• FsHandler log awareness: viewer commands (cat/head/tail/less/more/bat) now skip group_files_by_dir — the grouping step was collapsing non-path file content into a single ./ N modified summary, a latent bug that masqueraded as extreme compression on markdown/log fixtures. Viewer commands also switch to Keep::Tail when the target is tail or when a cat/less/more/bat path ends in .log/.out/.err — recent events beat initial noise for log viewing.
• Pre-existing test flakes (unrelated to the handler work): test_init_finalizes_prior_session_to_memory and the four structured-memory tests in test_memory_structured.rs hardcoded a 2026-03-22 session timestamp that fell outside the default 30-day retention window once wall-clock time drifted past 2026-04-21. Both suites now anchor on unix_now() - 86_400 and derive the date string from the same timestamp.

Motivation

Driven by an iOS development session where xcodebuild was the top bash command (~18 invocations) and squeez was routing it through the generic BuildHandler with only Gradle/Maven filters. Xcode-specific action noise survived compression unchanged — a clear gap for Swift/iOS workloads.

Benchmarks

• New fixture bench/fixtures/xcodebuild_build.txt: 1,881 tk → 17 tk (~99%)
• All 15 filter-mode fixtures pass.
• Efficiency proof: all 5 floors still pass. sig_mode_delta_vs_pipeline floor lowered 30% → 10% with a comment explaining the adjustment — the previous floor was measured against the grouping bug above.
• Aggregate benchmark: 91.6% reduction (152,961 tk → 12,886 tk), within normal variance of the prior 91.9%. README block refreshed.

Test plan

• cargo test — 159 lib + integration tests pass
• bash bench/run.sh — 15/15 fixtures pass
• ./target/release/squeez benchmark — 23 scenarios, no quality regressions
• ./target/release/squeez benchmark --efficiency-proof — all 5 floors pass