fix(lsp): charge each language server its own measured footprint (SCA-4688)

[pai-scaffolde]

Defect

The fleet cap derived from a single whole-fleet footprint constant, so every
language-server type was charged the typescript median. On the 16 GiB host
this fleet runs on:

budget                     = 25% of 16 GiB = 4096 MiB
LSP_CLIENT_FOOTPRINT_BYTES = 1300 MiB
derived cap                = 4096 // 1300 = 3

Measured healthy peak concurrency is 7 (SCA-4584, 11.6h span, 87 production
reap events). So the fleet evicted and respawned during ordinary multi-worktree
work.

Raising the constant was not available: 7 servers at 1.3 GiB is 9.1 GiB, which a
25%-of-RAM budget cannot express at any correct scalar value. The bind was the
single value, not its size.

Measurement (acceptance criterion 1)

scripts/measure_lsp_footprint.py (added here) spawns a real server against a
real project and sums physical footprint over the process subtree. RSS is
unusable: the Node-based servers run behind a shim reading 5-11 MB while the
analysis happens in a child.

server_id	measured	charged
typescript	1688.4 MiB (8.5k-file TS project)	1700 MiB
pyright	633.7 MiB	800 MiB
bash-language-server	49.1 MiB	150 MiB
yaml-language-server	41.2 MiB	150 MiB
anything else	unmeasured	1300 MiB (unchanged)

Note the direction: typescript was under-charged by the old constant, while
yaml was over-charged 31x. Raw readings: docs/lsp-footprint-measurements.json.
Every charged value is its measurement rounded up — headroom always makes
the cap stricter, never looser.

Change

The budget does not move. It is the same 25% of host RAM; servers fit better
only because most of them are genuinely cheap.

Admission is now two bounds:

• the existing count cap, now derived from the cheapest server rather than
  the most expensive one (leaving it sized off typescript would have kept the
  cap of 3 this fixes);
• a byte budget charging each live, spawning, and still-draining server its
  own footprint, reporting how many LRU victims the sweep must remove — the same
  unit _enforce_cap_async already loops on.

Acceptance criteria

1. Per-server_id values from a real measurement — table above, harness and raw JSON committed.
2. 16 GiB host admits the measured working set of 7 — test_mixed_fleet_of_seven_fits_the_incident_host, with test_the_same_fleet_would_have_been_over_the_old_flat_charge guarding the premise.
3. Typescript-heavy fleet still binds — test_typescript_heavy_fleet_still_binds, test_typescript_fleet_is_bounded_below_the_incident_footprint.
4. Unknown type falls back conservatively — test_unknown_server_charges_the_conservative_default.

Notes

• A single server larger than the whole budget (typescript on a small host) is
  never evicted — otherwise every spawn would immediately undo itself.
• _shutting_down is keyed by task with the client key as the value; charging
  the task would have silently mis-charged. Covered by
  test_shutting_down_servers_still_hold_their_memory.
• Stacked on fix(lsp): keep a wedged eviction off the caller's diagnostics budget (SCA-4628) #66. No branch in fix(tests): stop the LSP suite writing into the operator's real agent.log #61-fix(lsp): keep a wedged eviction off the caller's diagnostics budget (SCA-4628) #66 is modified, so their green CI runs and
  pending approval gates are untouched.

Verification

117 passed in 12.76s   (tests/agent/lsp/, includes the 12 added here)

Closes SCA-4688.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8bde1e1293

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".