fix(ui): stop syntect lazy-regex panics from aborting the app mid-render

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code

[Tryanks]

Root cause

The app crashed in the field with an instant SIGABRT while switching to a running child thread (crash report tcode-2026-07-28-044740.ips, v0.1.33). Binary forensics on the stripped release binary recovered the chain:

1. assets/syntaxes.bin comes from bat v0.26.1, whose syntax set is pre-tested against Oniguruma — but this crate compiles syntect 5.3.0 with the fancy-regex backend. 21 patterns (JS Babel \g<-1> subroutine calls, variable-length look-behind, \o{} escapes, ...) cannot compile under fancy-regex.
2. syntect compiles grammar regexes lazily on first use and panics with regex string should be pre-tested (regex.rs:70) the first time content exercises such a rule. One ordinary jsx arrow function is enough — reproduced: const f = async (a, b) => a + b; panics through tcode's exact highlight path.
3. The panic fired mid-GPUI-draw; unwinding through the element tree escalated into a double panic ("panic in a destructor during cleanup") → abort() → instant exit with no log. Lazy per-rule compilation plus the highlight cache is why it looked non-deterministic.

Fix (defense in depth)

• Data: regenerate syntaxes.bin via the new examples/sanitize_syntaxes.rs — patterns the shipped backend cannot compile are replaced with never-matching (?!) (22 replacements; 220 syntaxes and all 22,300 patterns preserved). Capture back-reference patterns are validated the way syntect's own YAML loader does (placeholder substitution before try_compile); all 44 such patterns pass.
• Runtime: highlight_source wraps syntect in catch_unwind, falls back to unhighlighted text, and poisons the offending syntax for the process (fail-closed on mutex poisoning) so a streaming code block cannot re-panic every frame.
• Latent sibling bug: the chat list's entry_range slice into the live timeline is now guarded with .get() — the snapshot can trail the timeline by a frame when adopting a running background thread.
• Regression tests: shipped_syntax_set_compiles_under_active_regex_backend fails CI whenever a future dump contains patterns the active backend rejects; the jsx repro test calls the unguarded parser directly so the fallback cannot mask a dirty dump.

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked — all green locally (macOS arm64).
• Repro snippet panics on the old dump, clean on the new one; sanitizer output re-verified by reloading the exact shipped bytes.
• Independently reviewed; review findings (fail-open poison lock, capture-pattern validation gap, fallback-masked test) addressed in the second commit.

🤖 Generated with Claude Code