fix(find): resolve hyphenated/suffixed GPU spec strings (Modal convention)

[scttfrdmn]

Summary

Modal's documented GPU spec-string convention hyphenates multi-word card names and appends a trailing !/+ suffix — gpu="RTX-PRO-6000", gpu="A100-80GB", gpu="H100!", gpu="B200+". pkg/find's tokenizer only splits on whitespace, so each of these arrived as one token and matched no vocabulary entry. Confirmed both by the original report and its follow-up comment (the bug is general — not narrow to RTX-PRO — since whitespace-only tokenization is the actual root cause).

Two independent bugs, both fixed:

1. pkg/find tokenizer: classifyTokens now falls back to normalizeHyphenatedToken for a word that failed every other single-word classification. It splits on hyphens and strips a trailing !/+, then re-runs classifyTokens on the resulting parts as if they'd been space-separated all along — so "rtx-pro-6000" resolves via the same matchPhrase alias lookup as "rtx pro 6000", and "a100-80gb" resolves as two tokens (GPU + memory) like "a100 80gb" already does. Guarded against infinite recursion (only reports "try splitting" when the word actually contains a hyphen or the suffix).

2. CLI routing (cmd/find.go's looksLikePattern): even with (1), real truffle find invocations still failed, because looksLikePattern decides whether to route to the pattern matcher or the natural-language parser before ParseQuery ever runs. Two bugs there: + is one of looksLikeRegex's own indicator characters, checked before the vocabulary check, so "B200+" never reached vocabulary recognition at all; and the vocabulary check compared the raw (non-lowercased) query against lowercase-keyed metadata tables, which was masked for most single-word terms by an accidental fallback but not for a hyphenated query like "A100-80GB". Replaced the old isKnownVocabularyTerm with a new exported find.IsRecognizedTerm (lowercases, then classifies via the same tokenizer, including the new hyphen/suffix fallback) and moved the vocabulary check to run first, before any regex-indicator check. This also fixed a latent case-sensitivity bug: uppercase vocabulary like "AVX2" previously only avoided misrouting by accident.

Verification against live AWS

Ran truffle find end-to-end (AWS_PROFILE=spore-host-dev) for every string named in the issue and its follow-up comment — all now return the expected instance types instead of "No matching instance types found":

truffle find "A100-80GB"    → p4d.24xlarge, p4de.24xlarge
truffle find "RTX-PRO-6000" → g7e.*
truffle find "H100!"        → p5.48xlarge
truffle find "B200+"        → p6-b200.48xlarge, p6e-gb200.36xlarge
truffle find "a100-80gb"    → p4d.24xlarge, p4de.24xlarge (lowercase)
truffle find "rtx-pro-4500" → g7.*

Test plan

• New pkg/find/hyphenated_test.go: TestResolveCard_HyphenatedSpecStrings (hyphenated form resolves to the same GPU/instance types as the equivalent space-separated form, for every case in the issue), TestResolveCard_HyphenatedFormResolves (full ResolveCard entry point), TestNormalizeHyphenatedToken (splitting helper, including the no-hyphen/no-suffix "don't split" guard), TestClassifyTokens_HyphenatedNoInfiniteRecursion (an unrecognized hyphenated word terminates as 3 TokenUnknown parts, doesn't loop), TestIsRecognizedTerm (the new exported helper, including the uppercase case that previously only worked by accident).
• New cmd/helpers_test.go: TestLooksLikePattern_HyphenatedGPUSpecStrings — asserts every case above routes to the NL parser, not the pattern matcher, and that a real hyphenated instance family (m7i-flex) still correctly routes as a pattern (the fix doesn't swallow every hyphen).
• Confirmed regression: stashing cmd/find.go + pkg/find/parser.go makes go vet fail to compile the new test file (undefined: normalizeHyphenatedToken) — the tests cannot pass without the fix.
• go build ./..., go vet ./..., go test ./... all green
• golangci-lint run ./... — only pre-existing findings in untouched files (app.go, gendocs.go, printer.go, root.go), none in pkg/find or the touched cmd/find.go/cmd/helpers_test.go
• make check-docs clean (regenerated docs-gen/find.md for the updated --help text documenting hyphenated spec-string support)
• CHANGELOG.md updated under ## [Unreleased]

Fixes #130

[codecov]

Codecov Report

❌ Patch coverage is 95.65217% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/find/parser.go	95.00%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!