fix(httpsig): percent-encode non-token keys and body-part names

[timothynodes]

Summary

Two related fixes for AO-Core message keys whose names are not valid HTTP header field names (spaces, emoji, uppercase, etc.).

hb_escape.erl

Add is_http_token/1 and encode_http_key/1. A key is percent-encoded only when it is not already a valid lowercase HTTP token, so already-valid keys stay byte-identical and human-readable on the wire while everything else becomes a legal header name. The transformation is reversed by the existing decode/1.

dev_httpsig_conv.erl

• encode_ids/1 now percent-encodes any non-token key instead of only ID-shaped keys (which were matched by byte size). Previously a tag named e.g. my <emoji> tag was emitted verbatim as an illegal header and rejected downstream — observed as a 502 from a fronting proxy.
• encode_body_part/4 / from_body_part/3 now symmetrically percent-encode and decode the Content-Disposition name. Large (> MAX_HEADER_LENGTH) and/or nested values are lifted into their own body part keyed by their full flat path — a name encode_ids/1 never sees — so their raw bytes reached the wire and crashed the structured-field parser on decode (parse_string/2 only accepts 0x20–0x7E). This mirrors the existing percent-encoding of committed key names in dev_httpsig_siginfo. The / path separator is preserved so nested-path splitting still works.

Compatibility

Keys/part-names that are already valid (lowercase letters, digits, -_. and the / separator) are byte-unchanged, so existing signed messages and their IDs are unaffected. Only names that previously could not round-trip at all change form.

Tests

Round-trip tests added for both fixes; existing tests still pass.

• hb_escape: is_http_token_test, encode_http_key_test
• dev_httpsig_conv: encode_ids_round_trips_weird_keys_test, encode_body_part_escapes_weird_name_test, encode_large_nested_weird_key_round_trips_test (end-to-end httpsig round-trip of a >4096-byte nested weird-named value)

🤖 Generated with Claude Code

[timothynodes]

Here's the test description with a concise Expected result:

---

Test: httpsig non-token-key header encoding (502 regression)

Purpose
Verify /~cache@1.0/read no longer returns 502 when a cached item has a tag whose key name is not a valid HTTP token (space, emoji, uppercase). Exercises commit 097268a.

Setup

• hb.timothynode.com → node with the fix.
• hc.timothynode.com → node without the fix.
• Cache key: UB33qavPDZkjCQ0Ezs0TqIsJs3XVjPpyibumrKSL7F8 (contains non-token tag names).

Steps
curl -s -o /dev/null -w "%{http_code}\n" "https://hb.timothynode.com/~cache@1.0/read?read=UB33qavPDZkjCQ0Ezs0TqIsJs3XVjPpyibumrKSL7F8"

curl -s -o /dev/null -w "%{http_code}\n" "https://hc.timothynode.com/~cache@1.0/read?read=UB33qavPDZkjCQ0Ezs0TqIsJs3XVjPpyibumrKSL7F8"

Expected result

┌──────┬─────┬────────┐
│ Node │ Fix │ Status │
├──────┼─────┼────────┤
│ hb │ yes │ 200 │
├──────┼─────┼────────┤
│ hc │ no │ 502 │
└──────┴─────┴────────┘

Observed ✅ 200 (fixed) vs 502 (unfixed) — matches expectation.