Add scoped_regex class in regex.hpp that automatically calls regfree()
on destruction. Replace all manual regcomp/regfree pairs in
awk_executor.cpp (4 sites) and expr.cpp (1 site), eliminating
potential resource leaks on early return paths.

Add unique_pipe type with PipeCloser to ensure pclose() is always
called, even on early returns. Eliminates potential pipe leak in
command substitution $(...) handling.

…tion

Remove zlib dependency entirely. Implement deflate compression (fixed
Huffman + LZ77 with hash chain matching) and inflate decompression
(fixed/dynamic/stored block support) in ~600 lines of header-only C++23.

Changes:
- New deflate.hpp: BitWriter with MSB-first Huffman encoding, LZ77 matcher
- New inflate.hpp: BitReader with accumulator-based peek/read, Huffman
  table builder supporting fixed and dynamic tables
- Rewrite compress.hpp: gzip format (RFC 1952) using our own deflate
- Update unzip.cpp: use raw_inflate instead of zlib
- Remove zlib from CMakeLists.txt (no more CPM fetch)
- Add 15 compression tests (round-trip, edge cases, corruption)

Verified: output compatible with system gunzip, all 331 tests pass.

Add -fvisibility=hidden, -ffunction-sections, -fdata-sections,
-Wl,--gc-sections, -Wl,--strip-all, -Wl,--hash-style=gnu,
-Wl,--build-id=none to Release builds. Debug builds unchanged.
Release size-optimized binary: 523KB.

Replace the quadratic LCS DP table with Myers' shortest edit script
algorithm, reducing memory from O(mn) to O(n+m) and time from O(mn)
to O(nd) where d is the edit distance. Add proper multi-hunk support
for unified diff output with configurable context lines. Eliminates
the code duplication between lcs_diff and unified_diff.

std::regex is known to be 5-10x slower than POSIX regex. Replace with
regex_t via the existing scoped_regex RAII wrapper. Also removes the
<regex> header dependency, reducing compile time and binary size.

…d/constexpr annotations

- Add streaming for_each_line() to io.hpp, refactor stream.hpp to delegate
- Stream grep/cat/wc — no full-file load, handles infinite input
- Optimize sort with precomputed keys, add reserve() across applets
- Mark 72 functions noexcept/[[nodiscard]]/constexpr across 6 headers
- Update README/architecture/Roadmap with current stats and benchmarks