Test the registry lookup, register MRU, document case-insensitive names - #382

…sitive names

The registry this PR originally proposed landed in #324, so what is left is
the part around it. Rebased onto develop after the docs rewrite (#372) and
the LRUv0 fix (#373).

test_evictionAlgo gains a case for create_cache_using_plugin(). Nothing
exercised that entry point, which is the one the MINISIM profiler uses and the
one that failed in issue #304: it went straight to dlsym(), which cannot see a
built-in constructor in a statically linked build because the archive member
holding it is never pulled in. The case also pins the case-insensitive
matching, since MINISIM callers pass names in whatever case the user typed.

MRU and LRUv0 are compiled into the library but were absent from the
registry, so no CLI could reach them. LRUv0 was left out earlier because it
segfaulted on the first request; #373 fixed that, so it is registered now.
Both names are covered by the new test.

The SHARDS profiler compared both the profiler name and the algorithm with
strcmp(), so --algo=lru was rejected even though MINISIM and cachesim accept
it. It uses strcasecmp() now, which makes "names are case-insensitive" true
for every mrcProfiler path, as the docs now say.

Docs: the extend guide says names are matched case-insensitively and when an
algorithm also needs a branch in cache_init.h, the mrcProfiler option table
marks --algo as case-insensitive, scripts/README.md says the same for
--algos, and the README lists mru.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VD7L6jWd2YzyhfMJxMquc

MRU_evict advanced q_head and then wrote q_head->queue.prev unconditionally.
When the evicted object was the only one in the queue, the new head is NULL,
so the second miss in a one-object cache (or any eviction that empties the
cache, which variable-sized traces can need) segfaulted. q_tail was also left
pointing at the freed object.

Handle the empty-queue case the way LRU_evict does: clear q_tail when the
head runs out. Registering mru in the previous commit is what makes this
path reachable from cachesim and the MRC profiler, so fix it here.

The new test drives a one-object MRU cache through four misses; it segfaults
without the fix and passes with it, including under ASan and LeakSanitizer.
Miss ratios on the sample trace are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VD7L6jWd2YzyhfMJxMquc

LRUv0_get inserted every miss itself instead of calling cache_get_base, so
it never asked can_insert. With lruv0 now reachable from the CLI, -a was
silently ignored: on cloudPhysicsIO at 0.1 with -a size and size=4096, LRU's
miss ratio is 0.9258 while LRUv0 stayed at its no-admission 0.8051. The
custom get also skipped the prefetcher and never counted requests.

LRUv0 already sets find, insert and evict, which are what cache_get_base
calls, so LRUv0_get now just calls it, as LRU does. With size admission
LRUv0 now matches LRU at 0.9258, its miss ratios without admission are
unchanged, and the existing LRUv0 test still matches LRU's expected results.

The new test attaches a size admissioner and checks that an object over the
threshold is never cached; it fails without the fix and passes with it,
including under ASan and LeakSanitizer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VD7L6jWd2YzyhfMJxMquc

The previous commit routed LRUv0_get through cache_get_base, which only
evicts on a miss. LRUv0_find, unlike LRU's, adopts the larger size when a
hit carries one, and the old LRUv0_get then evicted until the cache was back
under its size. Without that loop a 100-byte cache given (1,60), (2,40),
(1,80) held 120 bytes and still hit on object 2.

Keep cache_get_base for admission, prefetching and request counting, and run
the eviction loop after it again. LRUv0's miss ratios on cloudPhysicsIO now
match develop's exactly at 0.01, 0.05, 0.1, 0.3 and 0.6 without admission,
and it still honors -a.

The new test replays that sequence; it fails on the previous commit and
passes now, including under ASan and LeakSanitizer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VD7L6jWd2YzyhfMJxMquc

Registering lruv0 made the CLI hand it options its implementation never
sees. LRUv0 is the older GLib-based LRU and keeps its own hash table, so it
bypasses cache_find_base, cache_insert_base and cache_evict_base, and with
them everything built on those paths: admission (45fc5e1 patched that),
TTL expiry and default TTLs, and the prefetchers' find and evict callbacks,
so OBL never turns prefetching on and PG never builds its graph. Each review
round found another of these, and patching them one at a time does not
converge; doing it properly means moving LRUv0 onto the shared paths, which
is its own change.

So go back to what the original PR did and leave lruv0 out, with a comment
that gives the actual reason now that the crash is fixed (#373). Revert
45fc5e1 and 4ffe9ff, which only existed to make the registered lruv0 behave,
so LRUv0.c is untouched by this branch. The lookup test now checks that
lruv0 stays unregistered.

This reverts commits 45fc5e1 and 4ffe9ff.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VD7L6jWd2YzyhfMJxMquc