fix: make ctx smart deterministic and semantic-first

[saldestechnology]

Problem

ctx smart "<task>" selected the wrong files and was non-deterministic. For "add a new output format to ctx sql" it returned generic hubs (main.rs, smart_cmd.rs, analysis.rs) and even unrelated test files while omitting src/commands/sql.rs — despite that file's symbols (run_sql, render) being the top ctx semantic hits. Back-to-back runs of the same query produced different file sets (sql.rs appeared in only 2 of 5 runs).

Root cause

Two defects in rank_files (src/smart.rs):

1. Scale inversion. A direct SemanticMatch is scored with the raw similarity, which 1/(1+L2) (src/embeddings/mod.rs) compresses to ~0.4–0.6, while every depth-1 call-graph neighbor gets a flat 0.8 (CalledBy) / 0.7 (Calls). So a file pulled in purely by call-graph expansion outranked the on-topic seed that produced it, then consumed the 8000-token budget first (select_by_token_budget is greedy first-fit) and evicted the seed.
2. Non-determinism. rank_files sorted by relevance_score alone with no tie-break, over a Vec collected from a HashMap. The many files tied at 0.8/0.7/0.5 came out in hash order, so which tied files survived the budget varied per run.

(The underscore/ESCAPE bug from #15 is unrelated — smart resolves the graph via id-based call_graph/impact_analysis, never find_symbols_filtered.)

Fix

Rework rank_files into a tiered, total ordering:

1. Files with a direct semantic match rank above graph-only files (new FileSelection::best_semantic_score), so call-graph expansion supplements the seeds instead of displacing them.
2. Within a tier, higher score first — semantic score for the semantic tier, relevance_score for the graph-only tier; the two scales are never compared against each other.
3. Ties break by path ascending → deterministic across runs.

Seeding, call-graph expansion, weight() constants, the add_reason max-merge, and the token budget are unchanged — tiering supersedes the cross-scale comparison, so this is a ranking-only change. (A future overhaul could normalize the semantic/graph scales and add a lexical path boost; noted, not done here.)

Tests

The existing ranking tests only used single-tier Explicit reasons, so they never exercised the semantic-vs-graph inversion. Added:

• test_semantic_match_ranks_above_graph_only — a SemanticMatch{0.5} file must outrank a CalledBy{depth:1} (0.8) file.
• test_semantic_tier_orders_by_score_then_path — intra-tier score order, path tie-break.
• test_rank_files_is_deterministic — same set in any input order → identical ranked paths.

Verification

• Full suite green: 381 tests, 0 failures; cargo fmt --check and cargo clippy --all-targets -D warnings clean.
• End-to-end against a real index: ctx smart "add a new output format to ctx sql" run 5× now returns the identical set every time, with src/commands/sql.rs at rank 1. Spot-checks for "parse solidity contracts" and "generate embeddings with openai" surface the relevant implementation files (parser/mod.rs, embeddings/mod.rs) and are likewise deterministic.