Maximize frame throughput across all effects

[dhh]

This PR completes the FPS autoresearch pass with eleven separately measured optimization commits.

Combined result

Pinned-core, alternating-order benchmark against master (ca54938), using 11 samples per effect on the established 100x30 corpus:

• Aggregate FPS across all 35 non-clock effects: 19,873.98 -> 23,305.04 (+17.26%)
• Geometric-mean per-effect speedup: 1.185x
• Median per-effect speedup: 1.178x
• Total median runtime: 1299.337 -> 1108.044 ms (-14.72%)
• Faster effects: 35 of 35
• Complete frame streams: byte-identical for 35 of 35 effects

A dense 200x50 comparison across Beams, Waves, Colorshift, Highlight, Expand, and Middleout improved aggregate FPS by 32.49%. All six complete streams remained byte-identical.

Accepted wins

• Render directly from the reusable cell buffer, removing a full-frame copy.
• Evaluate quadratic Bezier paths without per-frame heap allocation.
• Parse the xterm palette once and memoize RGB conversions. Sampled xterm workloads improved 3.17x to 6.61x.
• Initialize Beams character scenes once. Standard FPS improved 2.151x in the final combined benchmark; dense FPS improved 1.985x.
• Skip unchanged Matrix rain visuals while preserving RNG draws. Isolated 200x50 runs improved 1.86x to 5.68x depending on color mode.
• Borrow SequenceEaser deltas instead of cloning cumulative character sets.
• Reuse the Overflow queue and borrow gradient state.
• Abort on panic in release builds. The final release binary is about 31% smaller than the original master binary.
• Store RGB colors inline with cached parsed channels. This alone delivered +11.6% geometric-mean FPS across 35 effects.
• Use an adaptive sorted-vector/bitmap active-character set. Dense effects improved roughly 6% to 14%.
• Bucket character groups in one pass. The grouping primitive improved 7.03x overall, with Matrix first-frame build improving 5.2%.

Validation

• cargo test --release: 23 tests passed across unit and integration suites
• Effect parity: 354 passed, 0 failed
• Terminal byte-stream parity: 41 passed, 0 failed
• CLI corpus: 19 passed, 0 failed
• GitHub CI: Linux tests/parity, macOS tests, and static musl binary all passed
• git diff --check: clean

Compatibility note

The compact RGB representation changes the public Rust field types from heap-backed strings to RgbString and prevents external Color struct literals. The project is binary-focused; command behavior and rendered output remain byte-identical.

[dhh]

Two more measured wins landed:

• aba5ead: stack-only quadratic Bezier motion. Paired isolated results: Fireworks +5.74%, Laser Etch +1.95%. Quadratic-effect parity: 40/40.
• 9238455: shared xterm conversion cache plus one-time palette parsing. At 200x50: Waves 4.03x, Sweep 3.17x, Overflow 6.61x faster. Xterm parity: 70/70; reference comparison: 4,913 RGB samples.

[dhh]

Two effect-specific wins landed:

• 43c22fb: initialize Beams scenes once instead of once per row and again per column. Default +72.5% FPS; dense 200x50 +71.7%; parity 8/8.
• 394109d: skip rebuilding unchanged Matrix rain visuals while preserving RNG draws. 200x50/600 frames: RGB 2.81x, no-color 1.86x, xterm 5.68x; parity 12/12 and full-stream hashes identical.

[dhh]

de78539 removes SequenceEaser deep clones by returning borrowed added/removed deltas. Dense 200x50 marginal frame time: Wipe -2.2%, Highlight -14.9%, Sweep -7.6%. Focused parity 30/30; full tests pass.

[dhh]

6b98670 removes Overflow gradient deep-copying and switches the pending-row queue from O(n) front removals to VecDeque. End-to-end gains are +0.6% to +1.4% across RGB/xterm/no-color and 64x20/200x50; parity 12/12 and full tests pass.

[dhh]

8efa1f3 enables panic=abort for release builds: paired all-35 throughput +1.38%, binary size 3.28 MB -> 2.55 MB (-22%). Release tests, 354 effect parity cases, and 41 TTY parity cases pass.

[dhh]

539723a stores RGB spellings/channels inline so Color is Copy: Color shrinks 56 -> 22 bytes and ColorPair 112 -> 44 bytes. Across 35 non-clock effects: +11.6% geometric-mean FPS, +12.0% median; dense 200x50 Beams +18.6%, Waves +18.2%, Smoke +16.8%. Candidate/base streams were identical for 354/354 effects and 41/41 TTY cases; release tests and CLI pass. Note: this intentionally changes the public rgb_color field from String to RgbString.