refactor: remove duplicated logic and parse each token once

[bsg62]

Quality-only pass over the Go sources: reuse, simplification, efficiency, and altitude. No user-facing behavior changes except the one noted below. 624 insertions, 968 deletions.

Production code

Parse once per run. decodeJWTHuman parsed the token, verifySignature parsed it again, and verifyClaims a third time, so a single --key --verify-claims run decoded the same segments four times and built eight maps to show three. parseUnverifiedJWT now returns a parsedJWT (raw string, token, segments, claims) threaded through the decode, signature, and claim steps.

The re-parse existed because formatTimestamps rewrites claim values in place, leaving the parsed map fit only for printing. printParsedJWT now formats a maps.Clone, so the parse stays authoritative and reusable.

One authority for key precedence. classifyKeyArg described itself as mirroring loadKeyForKID's ladder, but a mirror is a second copy. loadKeyForKID now switches on classifyKeyArg and delegates to loadSymmetricKeyFile/loadKeyFile/loadInlineKey, so the reading the CLI reports is by construction the reading that happens.

Deduplication.

• parseAnyDER replaces the six-parser DER cascade that was written out verbatim twice.
• parseKeyData's SSH branch was unreachable — every caller discarded its error and rebuilt the message via unsupportedKeyError, which does its own SSH detection. Removed, so the hint exists once.
• partSize formats base64URLLen's result instead of re-decoding; jweEncryptedParts gives both output paths one splitter.
• printVerdict renders the VALID/INVALID line for the signature and claim checks, which had it in duplicate. The two verdicts stay independent; only the rendering is shared.
• Decrypted payloads decode once into any and type-switch, instead of an object parse followed by an array parse; printSection widened to any so the array case stops re-implementing it.
• isJWT next to isJWE, so token-shape dispatch has one definition per form.
• Fast path in the escape helpers, so unescaped output allocates nothing.

Tests

Suite time 4.2s → 1.7s. generateRSAKey generated a fresh 2048-bit key ~60 times per run; it is now generated once via sync.OnceValues and shared, with generateDistinctRSAKey for the ten sites that need a second key that must differ.

• Six near-identical "unparseable key must not become an HMAC secret" tests → one 16-row table, every input kept.
• Seven JWE algorithm-family tests → shared assertJWEHeaderOnly/assertJWEDecrypts; same 62 subtests over the same algorithms and payloads. jwe_test.go drops 434 lines.
• Five key-file writers and twelve inline filepath.Join(t.TempDir(), …) sites → writeTempFile/writePEMFile/writePublicKeyFile/writeJWKFile.
• Cross-file fixtures pinTime and decodeJWTJSONMap moved into helpers_test.go per AGENTS.md.

Behavior change

An existing but unreadable key file now reports reading key file "x": permission denied instead of falling through to a base64 attempt on its own path and reporting "neither a valid file path nor base64-encoded data". No test pinned the old behavior; documented in AGENTS.md.

Verification

gofmt, go vet, and go test -count=1 ./... clean. Beyond the suite:

• Consolidated tests confirmed not to be silently passing: 62 JWE subtests before and after, and an injected bogus expectation fails as it should.
• printEncryptedParts column alignment pinned against a literal (the existing test only checks substrings).
• Built binary smoke-tested for human output, --json, wrong-key exit 1, and the new unreadable-file error.
• /code-review over main..HEAD found no correctness bugs. It independently fuzzed the escapeFormattedJSONControls fast path against the old rune loop over 20k random byte strings including invalid UTF-8 (byte-identical), and checked all 60 generateRSAKey sites for the distinct-key requirement. Its two low findings are fixed in a66b0da.

Not done

The json.Valid guard in formatTimestamps was flagged as dead code by two reviewers. It is not — a test feeds hand-constructed json.Number("1/2"), "0x10", "1p2", which big.Rat.SetString accepts. Left in place with a comment explaining why.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q26LwBBSBPUpbf1kd4Ep3z