Large static instruction/context blocks are re-sent per turn instead of being cached or referenced

[baron]

Impact

Re-injecting large unchanged instruction/context blocks each turn inflates every provider request. Worst sessions showed prompt/tool-history injections up to about 1.75M characters.

Actual behavior

Agent turn assembly appears to re-materialize large static blocks, such as instructions and workspace context, per turn rather than separating stable/static segments from dynamic turn content.

Expected behavior

• Static instruction/context segments are byte-stable across turns so provider prompt caching can apply where supported.
• Large context artifacts are passed by reference where supported, instead of repeatedly inlined.
• Turn assembly distinguishes static and dynamic segments and logs their sizes for diagnostics.

Publishable evidence

• Worst sessions showed injections up to about 1.75M characters.
• AgentSession JSON totaled about 64 MB across about 2,075 files.

Scope / non-goals

In scope:

• prompt assembly segmentation,
• static/dynamic byte accounting,
• provider-cache-friendly stable segments,
• diagnostics.

Out of scope:

• changing instruction content,
• tool-history compaction, tracked separately in Transcript compaction is turn-based, not byte-based; current-turn tool storms bypass compaction entirely #398.

Likely affected areas

• Sources/RepoPrompt/Features/AgentMode/Runtime/
• Sources/RepoPrompt/Features/AgentMode/Runtime/Claude/
• Sources/RepoPrompt/Features/AgentMode/Runtime/Codex/
• Sources/RepoPrompt/Infrastructure/AI/Prompts/
• Tests/RepoPromptTests/AgentMode/

Acceptance criteria

• Static prompt segments are byte-identical across consecutive turns of the same session when inputs do not change.
• Per-turn diagnostics report static versus dynamic payload bytes.
• Large artifact reference-vs-inline behavior is documented.
• Existing provider payload tests continue to pass.

Interlinks

Part of #406. Related: #398.

---