fix(watcher): stop endless refresh inside a linked worktree

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.

[BenLocal]

• Narrow the shared-gitdir watcher from worktrees/** to worktrees/*/{HEAD,gitdir,locked}: the broad glob also matched every worktree's private runtime state, and for a linked worktree that state IS our own gitdir. Every refresh runs git status, which takes index.lock, which classified as 'unknown' and forced another full refresh, which ran git status again — observed on a large monorepo worktree as a lock create/delete every ~1.5s (debounce + cooldown) that never stopped. Since any non-'status' change invalidates the history search, an in-flight search restarted from scratch each round and never returned a result. The three entries git worktree list actually reports are the only ones worth watching; from a main repo this also stops a sibling worktree's routine git activity from forcing a full refresh here.
• Keep the pattern deliberately "complex": per the VS Code API contract a RelativePattern containing path segments still creates a recursive watcher, so this only tightens the filter — it does not downgrade to a non-recursive watcher that would miss worktrees/<name>/HEAD entirely.
• Pin the registered patterns in a test: the failure mode is silent (no error, just a UI that never settles), so a regression would otherwise only surface as a bug report.