fix: prevent quadratic cache clears and preserve fork text styles

[zxch3n]

Local tree moves interleaved with deleted-container queries scanned the cache's peak table capacity on every move. Forks also lost the default text style, so marking a previously unused custom key failed with StyleConfigMissing.

Summary and rationale

• Replace the revivable deletion set with a fresh empty set on invalidation, including full cache clears. Entries have no destructor, so releasing the table avoids scanning its retained capacity. Keep final-deletion caching and all revival invalidation points intact.
• Copy the complete text style configuration in set_config, including its default. Both fork and fork_at use this path. Preserve the public config_text_style setter's map-only behavior.
• Add an ignored release scaling regression with three repeats and a median time(4n)/time(n) < 8 assertion. Add fork regressions for unused keys, explicit/default expansion, independent config copies, historical/detached forks, and the internal Rust API used directly by WASM.
• Add one loro-crdt patch changeset per bug and update context/dead-container-cache.md. One commit per bug.

Validation

Baseline: c00c9fa501f8d32f68d6255eacb7035a67fb6ab6 (origin/main, loro-crdt 1.16.4). The new scaling test failed before the cache fix (ratio 11.152), and all three fork regressions failed before the style fix with StyleConfigMissing.

• cargo test -p loro --test dead_container_cache: 11 passed; repeated with --release: 11 passed.
• cargo test -p loro-internal --lib dead_containers: 1 passed. The cache module's tests are gated by counter, so also ran cargo test -p loro-internal --lib --features counter dead_containers: 8 passed, including all 7 cache unit tests.
• cargo test -p loro --test fork_text_style: 3 passed; repeated with --release: 3 passed. Release cache tests also passed again after both fixes.
• CARGO_BUILD_JOBS=4 NEXTEST_TEST_THREADS=4 pnpm test: 1747 passed, 38 skipped; 62 doctests passed. Retried with network access after the clean worktree's dependency installation hit the sandbox's npm connection restriction.
• CARGO_BUILD_JOBS=4 pnpm check: passed (cargo clippy --all-features -- -Dwarnings, exit 0).
• Changed Rust files pass rustfmt --edition 2021 --check. The optional repository-wide cargo fmt --all -- --check reports existing formatting differences in unchanged files (loro_js_interop.rs, moon_transcode.rs, shallow_checkout_equal_value.rs, dag.rs, and undo.rs).

Release timing

Command: CARGO_BUILD_JOBS=4 cargo test --release -p loro --test dead_container_cache_perf -- --ignored --nocapture. Same workload before/after; three repeats, alternating 32k/128k order; creation, deletion, and cache priming are outside the timed loop. Other sessions share this machine, so scaling ratios are the primary result.

Version	Deleted nodes / moves	Repeat 1 (ms)	Repeat 2 (ms)	Repeat 3 (ms)	Median (ms)
Before	32,000	27.024	26.299	26.052	26.299
Before	128,000	293.281	292.806	293.800	293.281
After	32,000	7.094	7.627	8.039	7.627
After	128,000	28.697	30.041	32.189	30.041

Median scaling ratio: 11.152 before -> 3.939 after. The regression's bound is 8.

Moves-only controls:

Version	Moves	Repeat 1 (ms)	Repeat 2 (ms)	Repeat 3 (ms)
Before	32,000	6.572	6.314	6.429
Before	128,000	29.728	30.326	29.916
After	32,000	5.993	6.014	6.184
After	128,000	23.616	24.042	25.234

Control median scaling: 4.653 before, 3.998 after.

[github-actions]

WASM Size Report

• Original size: 3318.99 KB
• Gzipped size: 1104.60 KB
• Brotli size: 773.75 KB