build: separate build dirs for sanitizer and coverage builds

[battlesnake]

Problem

All sanitizer flavors write to the same platform build dir (packages/<pkg>/linux, windows, …). self.sanitize isn't part of the build path, so switching between asan / tsan / ubsan / none reuses one dir with a different sanitizer than it was configured for. run_config() then sees the sanitizer marker changed and forces a full CMake reconfigure every switch — slow, and occasionally a stale-mix build failure. Coverage (mama coverage) had the exact same problem: it landed in plain linux and collided with both plain and sanitizer builds.

Fix

Suffix the build dir per flavor via one build_dir_suffix() helper, reusing the existing CLI short tokens for sanitizers and adding a -coverage marker that composes cleanly in front of any sanitizer suffix:

flavor	dir
(none)	linux (unchanged)
coverage	linux-coverage
address	linux-asan
thread	linux-tsan
undefined	linux-ubsan
leak	linux-lsan
coverage + address	linux-coverage-asan

(and the same for every platform: windows-asan, android-tsan, linux-coverage, …)

The helper is applied in the two places that pick the per-platform name: platform_build_dir_name() and dependency_chain._save_mama_cmake's get_build_dir_defines (kept in sync per the existing WARNING comment, so dependency mama.cmake include paths match the suffixed dep dirs). No coverage and no sanitizer ⇒ no suffix ⇒ existing caches stay valid. Each flavor now keeps its own dir + CMakeCache.txt, so switching between already-built flavors no longer reconfigures.

Coverage

mama coverage build now lands in linux-coverage instead of sharing linux. The marker is prepended to the sanitizer suffix, so mama coverage asan build → linux-coverage-asan, keeping the previously-reviewed sanitizer dir names byte-identical (linux-asan, etc.). Plain mama build is unchanged (linux), so existing non-coverage caches stay valid.

Verification

tests/test_build_dir/ asserts the per-sanitizer names, the new coverage-only (linux-coverage) and coverage+sanitizer (linux-coverage-asan) names, and that no-coverage/no-sanitizer is unchanged. Full existing suite is green — 87 tests (prior 85 + 2 new coverage cases).

Ran the patched mama against a throwaway CMake project: mama coverage build, mama asan build, and mama coverage asan build create separate linux-coverage / linux-asan / linux-coverage-asan dirs; the MAMA_BUILD define in the generated mama.cmake matches the on-disk dir for each; re-running each already-built flavor skips CMake configure (no churn); plain build still uses linux.

Back-compat note

First build of each flavor after upgrading mama configures once into its new suffixed dir (the old shared linux dir keeps the plain no-coverage/no-sanitizer build). Expected, one-time.

Supersedes #22 (moved off fork to a branch on this repo).

LOC breakdown

	prod	tests
sanitizer dirs (8f07b02)	18	33
coverage dirs (this change)	3	13
branch total vs master	18	46

[RedFox20]

A comment got spliced

[battlesnake]

fixed, a clanker got spliced for that mistake

[RedFox20]

This is fishy and doesn't seem to belong here, the whole suffix definition should be done inside BuildConfig (single source of truth). This would also remove the unnecessary comment must match platform_build_dir_name() -- because in reality even Claude will not often read this properly.

build_dir = c.build_dir_with_suffix(build_dir)

This would delegate the single source of truth into BuildConfig class, and BuildConfig.platform_build_dir_name() can also use it

[battlesnake]

Concrete motivation to land this: it would've caught KrattWorks/rtpvideo#4 — an unaligned *reinterpret_cast<const uint16_t*>(ptr) read in BitStream that mama ubsan build silently ran without -fsanitize (it reused the non-sanitized linux dir, exactly the stale-mix described above). With per-flavor build dirs mama ubsan actually applies the flag, and that read then trips the alignment check.