detect_tag_changes.py: normalize RISC-V macros in the change report (#245)

[Bill Traynor (wmat)]

Part of #245 (item 2 of 3) — the highest-value item.

The tags backend stores raw AsciiDoc (#169), so the reviewer-facing change report compared markup-migrated text as different: mip vs csr:mip[], 11b vs 0b11, -- vs {endash}. On the real riscv-isa-manual reference regeneration (riscv/riscv-isa-manual#3249) this made 199 of 216 "Modified" tags (92%) markup-only false positives, and the tool exited non-zero — which consumers treat as a breaking normative change.

Change

Teach _strip_asciidoc_formatting the semantic macros so a pure markup migration reports as no change:

form	normalizes to
csr:NAME[]	NAME
csr:NAME[FIELD]	NAME.FIELD (field upper-cased)
csr:[FIELD] / csr::[FIELD]	FIELD
ext:NAME[]	NAME (capitalized: F, D, Sscofpmf, Zicsr)
insn:NAME[]	NAME

Also normalizes binary renotation (11b → 0b11) and en dash (-- → {endash}).

This is applied only to the review comparison, never as a freshness check — erasing formatting is correct for "did the normative text change" and wrong for "is this file byte up to date", the exact conflation that caused ref/ to drift. --strict still compares prose byte-for-byte and is unaffected.

Validation

Ran the old vs new normalizer against the two committed riscv-isa-manual reference revisions spanning the #3249 macro migration:

• Old: Added 1, Deleted 2, Modified 216 (matches Normative tag tooling: unsound --update-reference, normalizer misses RISC-V macros, tags.rb omits trailing newline #245 exactly)
• New: Added 1, Deleted 2, Modified 32

The remaining 32 are genuine normative edits (added clauses, table restructures) plus a few markup forms outside this migration's scope (insn: with operands, single-hyphen ranges) that are appropriate for a human to review. make test passes.

Note: the issue's lone ext:f[]→F example was misleading — multi-letter extension names render capitalize-first (Sscofpmf, not SSCOFPMF); handled accordingly.

🤖 Generated with Claude Code

[Bill Traynor (wmat)]

Validated against riscv-isa-manual via the manual's exact detect_tag_changes.py invocation on the real #3249 macro migration: the change report drops from Modified 216 to Modified 32. Full evidence: #245.

[Jordan Carlin (jordancarlin)]

Seems like this might be tricky to maintain. It would be better to parse this from the macro definition file or something like that. Fine for now though and a clear improvement even as is.

[Rafael Sene (rpsene)]

I like that normalization stays confined to the review report while --strict remains byte-exact. +1 to Jordan on eventually sourcing the macro list from the definitions rather than hardcoded regex.