refactor(chat): extract the Markdown residency policy behind a pure internal seam

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code

[Tryanks]

Architecture round, candidate C5.

Problem — the eviction policy inside chat/mod.rs (2,356 lines) was deterministic but interleaved with store observation, list callbacks, and rendering; its tests had to create a real visual window, draw, manipulate ListState, redraw, and peek at private residents through cfg(test) accessors — testing past the interface because the policy had no seam of its own.

Change — new chat/residency.rs (202 lines, pub(super)): decide(ResidencyInput) -> ResidencyDecisions consumes the visible turn window, per-turn markdown entries, streaming state, and pins (composer tail, selection participants, one-shot turn target) and returns build/evict sets; every policy constant (±8 build, ±24 evict hysteresis, tail pins, viewport fallback) is named and documented in one place. ChatView is now the adapter that gathers inputs and applies decisions to GPUI entities. chat/mod.rs sheds 170 lines.

Wins — tests hit one pure interface (no windows); policy constants stop being buried locals; locality: hysteresis bugs concentrate in the policy module.

Behavior unchanged: hysteresis, pins, selection rule, tail-follow, and palette turn-target rebuild semantics survive; the 240-turn regression numbers are identical (48 at tail / 78 mid-scroll). One end-to-end GPUI test kept to prove ChatView applies decisions.

Gates: fmt / clippy -D warnings / full workspace tests — green locally.

🤖 Generated with Claude Code