fix(renderer): §82's escape hatch has to survive canonical rendering

[rmichaelthomas]

The defect

source:     require total is "large"        (with `define large: is above 50` above it)
canonical:  require total is large
re-parses:  RequireNode → PredicateApplicationNode   (was: RequireNode → ConditionNode → QuotedString)

Same text out, different program back. Reported during the CommonGage port and left unfixed at the time because changing canonical output looked like a migration.

Why both halves were right

v31 §82 makes quoting the escape hatch for a value that collides with a predicate name, and the parser implements it exactly — parser.py:3498 tests for an UNKNOWN token, so a QUOTED_STRING never reaches the predicate branch. There is already a test for this (test_quoted_string_forces_equality_even_if_predicate_name_matches).

v2c §90 drops the quotes around a "safe single word", where safe means not reserved. A predicate name is never reserved — it is declared. _emit_string's safety test predates define and cannot see it.

Neither checkpoint saw the other, and what fell through the gap is a rendering that means something else.

The fix

It follows the arrangement v2d §96 already reached for composition-call arguments — name-resolvable in exactly the same way, and for which a QuotedString has always kept its quotes (renderer.py:264). The parser is the only stage holding the predicate table, so it marks the collision on the literal and the renderer honours the mark.

The flag is compare=False, matching this file's existing idiom for inert metadata (starting_date / until_date, two declarations up). Two "large" literals are the same value whether or not one must be written with quotes to stay one — so parse(render(ast)) == ast still means what it meant.

Scope

Narrow, and asserted as narrow: require status is "active" still renders require status is active when nothing named active is declared. A later widening has to be deliberate rather than accidental.

Corpus

Case added. This is the exact class the behavioural corpus exists for — the two programs are one character apart and no word-list gate or text comparison can distinguish them. The nodes field does.

Verification

1899 passed, 0 failed. The new test was written first and watched fail with assert 'require total is large' == 'require total is "large"'. grammar/ regenerated (source-line drift only — no rule or error-text changes).

One thing this does not fix, deliberately

The same class exists in a second position, and there it changes a value, not a node kind:

remember a number called total with 75
remember a string called label with "total"     → canonical: ... with total
show label      original: "total"     canonical: 75

BareWord does symbol-table fallback and QuotedString does not, so normalising the quotes away converts one node into the other. Unlike the predicate case, this is not a gap between two rules — §90 says single-word quotes drop by design, and three tests assert it by name (test_render_round_trip_single_word_quote_drops). Closing it means reversing a locked decision, which is a call for @rmichaelthomas, not a fix to smuggle into this PR. Measured cost if we do: 6 failures out of 1899, three of them genuine spec assertions and three regeneration artifacts.

🤖 Generated with Claude Code

https://claude.ai/code/session_01E6qcDj1e1dEYvc8jLnpGZy