fix(lexer): stop treating ]] and unbalanced [...] as special outside conditionals

[mpecan]

Closes #35

Root cause

Two independent bugs in the lexer / parser both triggered by ]] and [...] appearing outside a [[ ]] conditional.

Bug A — ]] was always a reserved word. TokenType::reserved_word("]]") returned DoubleRightBracket regardless of context, and Parser::is_list_terminator treated any Word "]]" as a list terminator. A stray ]] in the middle of a plain command (e.g. Cdeclare -n ref=t ]] arget) silently terminated the list instead of becoming an ordinary word.

Bug B — read_bracket_subscript was unguarded. The helper fired on any mid-word [ as long as the word so far was non-empty and didn't already end with [. That absorbed whitespace and control operators inside constructs like ][[, [c[, $$[a||b], case[a&&b], and foo^[...] where bash does not.

Fix

• src/lexer/words.rs — gate read_bracket_subscript so it only runs when:
  • we're at command-start and the word so far is a plain bash identifier (arr[i]=val assignments, foo[...] invocations, which bash parses as a single word and in which subscripts permissively absorb whitespace), or
  • we're inside [[ ]] (so regex character classes like [[:alpha:][:digit:]] still tokenize as one word).
• src/lexer/words.rs — after the existing reserved-word classification, downgrade TokenType::DoubleRightBracket to TokenType::Word whenever cond_expr is false. parse_cond_command sets cond_expr across the body of a [[ ]], so the reserved-word meaning is preserved exactly where bash expects it.
• src/parser/mod.rs — drop the Word "]]" branch from is_list_terminator. A bare ]] is now a plain word that flows into the current command instead of ending it.
• src/lexer/words.rs — remove the stale read_bracket_subscript call sites in continue_word and read_array_element (they are no longer needed once the main-loop branch is guarded).
• tests/integration.rs — remove the lexer: ]] not tokenized as separate word outside [[ ... ]] #35 entries and the bonus side-effect entries from KNOWN_ORACLE_FAILURES so they become active regression guards.

Fixed test cases from #35

• rbracket_outside_cond 1 — ][[ "$file" == *.txt ]]
• rbracket_outside_cond 2 — [c[ $x =~ ]+[a-z] ]]
• rbracket_outside_cond 3 — Cdeclare -n ref=t ]] arget

Side-effect fixes (tests, not auto-closing issues)

Side-effect fix for cases originally reported in #36 and #37 (case 5):

• bracket_op_split 1 — echo ho $$[a||b]
• bracket_op_split 2 — Decho $ case[a&&b]
• bracket_op_split 3 — foo^[a-echo ${foo^[a-z]}
• reserved_word_as_word 5 — if [ $"yes" = yes9 ][ $; then echo ok; fi

All four share the same root cause as Bug B and now pass. KNOWN_ORACLE_FAILURES has been updated and unit tests were added as direct regression guards. #36 and #37 are intentionally not auto-closed — #36 can be closed by whoever verifies there are no further manifestations, and #37 still has cases 1-4 open (parallel work).

Test plan

• cargo fmt
• cargo clippy --all-targets -- -D warnings — no warnings
• cargo test — 230 passed (175 unit + 1 doc + 48 integration + 6)
• 11 new unit tests in src/lexer/tests.rs exercise each failing scenario directly at the lexer level (no reliance on the oracle corpus)
• Regression guards in place for identifier-prefixed array subscripts (arr[0 foo]) and [[ ]] regex char classes ([[:alpha:][:dig||it:]])
• No regressions in the Parable corpus or the remaining oracle suites