AI: recover after losing part of its base

[CubeB]

From the ROADMAP (requested 2026-08-30). Today the computer player carries on down its build list after an attack, as if nothing had happened. It should:

1. Rebuild what was destroyed: remember each lost structure and its site, and requeue it ahead of new expansion.
2. Reclaim its own wreckage for the metal, especially when it is metal-starved.
3. Remember which direction attacks came from, and put defences on that side.
4. Hold ground sensibly: place units for defence and for the next push rather than leaving them where they stopped.
5. Scale with difficulty: all of this on Easy, Standard and Hard, with the thoroughness increasing with difficulty.

Where. src/rwe/ai/: BuildManager, EconomyManager, ArmyManager, StrategicManager, ThreatMap, PerceptionManager. See docs/ai-architecture-proposal.md. The AI runs inside GameSimulation::tick and emits ordinary PlayerCommands, so the determinism rules in CLAUDE.md apply: no uniform_int_distribution, use SimRandom.h's randomBelow, and any new state follows the save / hash / dump rule.

No original to copy. §103 could not establish that TA's computer player does any of this, so the design is RWE's own. Keep it in the AI's existing style, and write sim tests that build a base, destroy part of it, tick, and check what gets requeued, reclaimed and defended.

Suggest one PR per numbered point.

🤖 Generated with Claude Code

https://claude.ai/code/session_018moxVoyHDBoHqkw8uqS3HX

[HectorBailey]

Questions before the first of the five PRs:

1. Priority order. The list reads 1 rebuild, 2 reclaim wreckage, 3 remember attack directions, 4 hold ground, 5 difficulty scaling. Point 5 cuts across the other four, so I'd do it inside each rather than last. Of 1 to 4, is rebuild the one you most want first, or is it reclaiming its own wreckage (the metal-starved case), which is the smallest and touches only EconomyManager and the existing patrol reclaim?
2. What "remember" is allowed to be. Every one of these wants state the AI keeps between ticks (lost sites, attack directions), and the AI runs inside the deterministic sim, so that state goes through the save, hash and dump rule. Is it acceptable for the AI's memory to be hashed and saved (which makes an AI player's save file grow), or would you rather it be re-derived each tick from the sim's own state (the wreck list, the last-attacker fields the units already carry) so nothing new is hashed?
3. Difficulty semantics. AiDifficulty has Idle, Easy, Standard, Hard and Brutal. "Thoroughness increasing with difficulty": should Easy get none of the recovery behaviour (it plays as today), Standard rebuild and reclaim, Hard and Brutal all four? Or all tiers get all four with tuning knobs in AiTuningProfile, which is what --ai-tune already exposes?
4. A test fixture. The AI tests build a base and tick; none destroys part of one. I'd add a fixture that stands a base up, kills chosen structures with quietlyKillUnit or a blast, and ticks, and use it for all five. Any objection to that living in AiBehaviour.test.cpp next to the existing cases rather than in a new file?

[CubeB]

1. Rebuild first. It's the headline behaviour and was listed first when it was asked for. Then reclaim, attack directions, and holding ground. Agreed on folding difficulty into each rather than doing it last.
2. Re-derive where you can. Use the wreck list and the last-attacker fields the units already carry. For what can't be re-derived, such as the sites of lost buildings, keep it in the controller next to the AI's existing state. That state isn't saved or hashed today: aiControllers isn't in save_util or GameHash_util, and a loaded game builds its controllers fresh. Keeping AI memory across a save and load would be its own issue; don't take it on here.
3. Every tier except Idle gets all of it, scaled. The request was for it "on Easy, Standard and Hard, with the thoroughness scaling by difficulty". So Easy recovers too, just slowly and partially: longer delays, and only the economy rebuilt. Hard and Brutal are thorough. Put the knobs in AiTuningProfile so --ai-tune reaches them.
4. A fixture in AiBehaviour.test.cpp is fine.

🤖 Generated with Claude Code

https://claude.ai/code/session_018moxVoyHDBoHqkw8uqS3HX