fix: reconcile resumed tool name collisions

[seratch]

This pull request fixes stale tool and handoff routing when a run resumes after a human-in-the-loop approval.

Before this change, queued function-tool and handoff runs retained the tool objects selected before the pause. If the enabled tools, handoffs, or their ordering changed while the run was interrupted, resuming could execute a stale or disabled tool even when RunConfig.tool_name_collision_policy selected a different current winner.

The resume path now rebuilds the current tool inventory using a consistent enabled-handoff snapshot, applies the canonical collision resolver, and reconciles queued calls with the resulting winners before allowing side effects. Under the default warn policy, queued calls are rebound to the current winning function tool or handoff. Under the error policy, current collisions fail before tool execution or handoff transfer.

The implementation also:

• reconciles function-to-function, function-to-handoff, handoff-to-function, and handoff-to-handoff winner changes;
• respects tool_not_found_behavior when a queued tool is no longer available;
• preserves model call order when combining executed, rejected, and missing-tool results;
• validates persisted approval identity before dynamic tool inventory callbacks run;
• snapshots approval and invocation metadata so mutable callbacks cannot change routing or authorization decisions during resume;
• preserves interrupted Agent.as_tool() state when a strict missing-tool error or formatter cancellation aborts a resume;
• transfers nested run state only after reconciliation succeeds, preventing failed or ambiguous replacements from corrupting retryable state;
• propagates the same handoff snapshot through custom get_mcp_tools() and get_all_tools() overrides that delegate to the base implementation.

This does not add a new public API or change the serialized RunState schema. It uses the existing tool_name_collision_policy configuration.

Queued calls with duplicate call IDs intentionally retain the existing ordered owner-binding behavior because they cannot be reconciled unambiguously. Applications that need tool replacement across an interruption should use unique provider call IDs. Replacing an interrupted Agent.as_tool() with a different owner before deserializing the state also remains unsupported because the current persisted state does not contain a durable owner identity.

Documentation is intentionally excluded from this pull request and will be handled separately.

This pull request resolves #4116.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4517f74a93

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e0f0b501ad

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 69dddfc2d3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 521b12dedc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 96e096ef20

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".