BE hyperstack: add optimizer plugin + skill (teaches runtime analysis + algorithm derivation)

[KailasMahavarkar]

Adds the optimizer - hyperstack's missing quality axis. Code today is made current, correct, designed, and minimal; nothing made it algorithmically sound. This closes that, the hyperstack way: evidence over vibes, lens not gate.

What it is (and is NOT)

It does not ship a library of algorithm implementations - models already know BFS, and algorithms are stable (unlike framework APIs). Instead:

Part	Role
optimizer skill	teaches two transferable skills: (1) derive Big-O straight from code (cost model, recurrences/Master Theorem, amortized, space), (2) derive a better algorithm by removing redundant work ("what does this recompute that it could do once?")
optimizer_* plugin (5 tools)	the menu/lookup: problem -> candidate technique classes + complexity + the naive smell it replaces + a web-search query for the authoritative impl. ~60 techniques across 18 classes. No code shipped.

The skill is the thinking; the plugin confirms the named technique and points at the impl to web-search. Gated by YAGNI - it stays quiet when a naive loop over 10 items is correctly the lazy-right answer.

The derivation loop (the teaching)

derive current Big-O -> locate dominant cost -> name the redundancy ->
map redundancy to its fix (= technique class) -> optimizer_get_technique + web-search impl ->
re-derive new Big-O  (present before -> after)

Files / wiring

Area	Change
Plugin	src/plugins/optimizer/{data.ts, index.ts, tools/*} (13th plugin)
Skill	skills/optimizer/SKILL.md (teaches analysis + playbook)
Wiring	src/index.ts, bootstrap Layer-1 + Support-Skills, audit manifest (editorial), registry test 12->13

Verification

bun test 47/0 (registry now 13, audit coverage + skills-index include optimizer) · bun run build clean · bun run audit clean (optimizer = editorial) · 5 tools register and match_problem returns the menu.

Honest notes: (1) the matcher is substring-based, so it over-returns candidates (e.g. "unsorted" trips "sorted") - acceptable for a recall aid, worth tightening later. (2) No version/count bump here - that rides the next release (1.2.0), same as how 1.1.0 was cut.