Parser Escape Oracle

Purpose

This is the narrow corpus for the current compatibility problem.

It is not trying to model the whole evaluator. It only tracks the cases that distinguish:

TinyMUX 2.13 compatibility
current 2.14 AST behavior
Penn-style percent grammar

The corpus lives in parser/escape_oracle_cases.txt and is exercised by parser/run_escape_oracle.sh. That runner validates the production-shaped mux214 and penn study profiles by default. The exploratory mux213 study profile is optional because matching corrected deferred-region semantics there would require server-grade machinery the study tool does not fully implement.

For traced real-command behavior after command parsing/quoting, see the separate command-level oracle in docs/command-escape-oracle.md.

What Belongs Here

Only cases that answer one of these questions:

Does backslash stripping happen before or after percent dispatch?
Does a noeval path still consume backslashes?
Does brace-group selection change the answer?
Is the divergence caused by tokenizer grammar or evaluator behavior?

That means the oracle should stay focused on:

\% vs \\% vs \\\%
known % forms vs unknown % forms
plain evaluation vs if/switch/case/iter
brace vs non-brace noeval arguments
Penn-only grammar collisions such as % and %iL

Status Meanings

confirmed: checked against a running real engine and reflected in the corpus
needs_live: local model is useful, but still needs a live-server checksum

If a row turns out to be based on a bad translation or a stale study assumption, remove it or downgrade it. Do not keep known-bad rows in the oracle just because they are old.

The current 2.13 and Penn confirmations were captured from traced @pemit runs and recorded only in sanitized form. Player names and dbrefs are intentionally not preserved here.

Important: the oracle corpus is parser-level. If a traced command goes through additional quoting or command parsing first, that trace is still useful evidence, but it must be translated back to the effective eval input before being copied into the corpus. If that translation is not clear, record it in the command-level oracle instead.

Current Spec Split

There are two separate targets. They should not be merged into one mode.

Target A: TinyMUX 2.13 Compatibility

This is the regression-fix target for 2.14 AST work.

Requirements:

match 2.13 ordinary % and \ behavior
match 2.13 FN_NOEVAL multi-pass behavior
preserve 2.13 oddities even when they are awkward

This target is driven by live-engine checksums, not aesthetics.

Target B: Optional Saner Backslash Mode

This is a future feature target, not a compatibility fix.

Requirements:

explicit opt-in
documented as intentionally non-2.13
specified independently from the compatibility work

Penn can inform this mode, but Penn behavior should not be smuggled into the 2.13-compatibility target.

Current Cleanup

The focused oracle was recently pruned after some deferred-NOEVAL rows turned out to be wrong. In particular, the old claim that TinyMUX 2.13 lost the percent for [switch(1,1,{\\% capacity})] was false; that row has been revalidated live and corrected.

The current rule is simple:

keep rows only when their parser-level input is clear
correct rows when live validation disproves them
remove rows when the live basis is not trustworthy enough yet

That leaves a smaller oracle, but it is a cleaner specification target.

3.6 KiB Raw Permalink Blame History