tags.rb: substitute AsciiDoc attributes in tagged inline text

[Bill Traynor (wmat)]

Fixes #169.

Two commits: the attribute-substitution fix, and the removal of
detect_tag_changes.py per Jordan Carlin (@jordancarlin)'s review.

1. The problem

Asciidoctor applies substitutions at the block level, in this order:
specialcharacters, quotes, attributes, replacements, macros,
post_replacements. Inline nodes are produced by the quotes pass, so their
text has only had specialcharacters applied — an attribute reference such as
{ge} reached the tags JSON unresolved, and there is no built-in way to ask an
inline node for its substituted text.

PR #131 fixed that by replaying the parent block's whole sub list over the
inline text. That also ran the three passes after attributes, which is what
made it a net loss and got it reverted in PR #170:

• replacements rewrote apostrophes to ’
• [.underline] was dropped
• <<xref>> cross references were deleted outright

The fix applies only the attributes substitution. Nested formatting is
unaffected, because bold, italics, underline and friends are separate inline
nodes that reach the converter in their own right — they never pass through
this code path as raw text.

Verified against riscv-isa-manual

Built make build-tags on riscv-isa-manual at main with the submodule on
each revision. No warnings under --failure-level=WARN. Of 2923 tags, 38
change, all of them attribute expansions; the tag set is identical and
nothing else moves. Cross references survive (<<translation>> in
norm:sfence_vma_asid_only), apostrophes stay literal (two's, entry i's),
and superscripts and insn:/csr:/ext: macros are untouched:

norm:elen    ELEN {ge} 8       ->  ELEN ≥ 8
norm:vwsll   2{times}SEW       ->  2×SEW
norm:c-j_op  a {pm}2 KiB range ->  a ±2 KiB range

A note on escaping

{ge} expands to ≥ because symbols.adoc defines it as literal Unicode.
{endash}, {emdash} and {asterisk} are defined as character references
instead, so they expand to the &-escaped form — ‍–‍.
That looks like over-escaping but is the consistent result: every string this
backend emits sits at the post-specialcharacters stage (<<, &),
and an author-written ≥ already produces ≥ today — see
norm:unicode_decimal_value. Both spellings render identically through
create_normative_rules.py.

Tests

Three cases added to tests/norm-rule/test-ch1.adoc, pinning all three
regressions from #169 plus the attribute-missing default:

"norm:attribute_in_formatting":  "≥ and ≥"
"norm:attribute_with_hyperlink": "It isn't 100, it's ≥ <<norm:superscript>>"
"norm:attribute_undefined":      "200 {no-such-attribute} 100"

Attribute expands through nested formatting; apostrophe stays literal; cross
reference survives; an undefined attribute is left alone.

2. Deleting detect_tag_changes.py

riscv/riscv-isa-manual#3334 removed the only consumer — the
check-normative-tags workflow, the committed ref/riscv-spec-norm-tags.json
and the supporting scripts — because the reference copy existed only to raise
automated issues, and those are no longer wanted. Nothing else calls the
script: no importers, no workflow here, and the only remaining references were
this Makefile and the tools README.

Normative rules are still built in CI and shipped in every release. That path
goes through create_normative_rules.py and is untouched.

This also retires test-tag-changes-update, which has been failing since
PR #248 removed the --update-reference flag the target still passed.

Result

make test exits 0 from a clean build/ — 16 assertions, 0 failures, no make
errors, for the first time since #248.

🤖 Generated with Claude Code

[Bill Traynor (wmat)]

Jordan Carlin (@jordancarlin) would you mind checking this one? I think both James and Tim have moved on from this.

[Jordan Carlin (jordancarlin)]

I think long term we should think about simplifying a lot of this, but this seems fine for now.