chore(build): reduce micasa binary size (~54MB)

[cpcloud]

Problem

The micasa binary is ~54 MB stripped (go build -ldflags="-w -s" -trimpath ./cmd/micasa). That's heavy for an interactive TUI and bloats download archives and container layers.

On-disk breakdown (53.5 MB)

Section	Size	Notes
.text (machine code)	~22 MB
.gopclntab (PC→line tables)	~18 MB	needed for stack traces; scales ~linearly with code
.rodata (strings, embedded data, type info)	~12 MB
.data / .noptrdata / .typelink / .itablink	~2 MB
.bss	(33 MB)	not on disk — the crypto/internal/fips140/drbg.memory 32 MiB scratch buffer, runtime-only

Biggest discretionary contributors

1. Multi-provider LLM stack — internal/llm/client.go unconditionally imports 10 any-llm-go providers. The gemini provider drags in google.golang.org/genai → google.golang.org/grpc + google.golang.org/protobuf + cloud.google.com/go/auth (s2a-go, gax-go, …). anthropics/anthropic-sdk-go (~1.4 MB) and openai-go (~1.8 MB, shared by openai/groq/mistral/deepseek/llamacpp/llamafile) are also linked. Rough total incl. proportional gopclntab/rodata: ~12–16 MB.
2. alecthomas/chroma/v2 lexers via glamour markdown rendering — code + dozens of embedded language lexers. ~2–4 MB.
3. itchyny/gojq via internal/config/query.go (config query <jq> feature) — full jq engine + yacc parser tables. ~1.5 MB.
4. brianvoe/gofakeit/v7 via internal/fake (demo data) — code + embedded word lists. ~1.5 MB.

Quick wins already identified

• .goreleaser.yaml ldflags only sets -X main.version=… — no -s -w, so released archives aren't even stripped (~3–5 MB larger than they need to be).
• upx --best --lzma takes the stripped binary 53.5 MB → 12.5 MB (measured). Tradeoffs: small one-time decompress on launch, breaks go version -m, occasional Windows AV false positives, friction with macOS code-signing/notarization.

Options to weigh

• A. Add -s -w to goreleaser ldflags (trivial, no downside).
• B. Add a UPX post-build hook in goreleaser (4× smaller downloads; band-aid — same runtime memory + container size; macOS may need exclusion or post-pack signing).
• C. Collapse / replace the any-llm-go multi-provider stack with thin direct HTTP clients (OpenAI-compatible chat-completions + Anthropic Messages cover everything micasa does — NL→SQL and summary calls). Real fix, ~10–15 MB, breaking refactor of internal/llm. Aligns with "LLM is opt-in, not a crutch."
• C1 (lighter). Just drop the gemini provider import to shed genai/grpc/protobuf/cloud-auth (~6–10 MB), keep the rest.
• D. Drop gojq from config querying (replace with a path selector or just config show) — ~1.5 MB.
• E. Trim chroma lexers down to the few languages actually highlighted — needs a fork/vendored registry; brittle across chroma upgrades. ~2–3 MB.

Recommendation

Do A + B now for shipped artifacts (≈ 57 MB unstripped release → ≈ 12 MB), then C/C1 as a follow-up to cut the uncompressed footprint (UPX-decompressed memory and container layers still carry the full size).